The Taste Layer

Abstract

The rapid integration of generative artificial intelligence into marketing content production has triggered two parallel and seemingly contradictory developments in the digital marketing landscape. On one hand, peer-reviewed field studies demonstrate that AI-generated visual marketing can match and frequently exceed the click-through performance of human-made advertising (Hartmann et al., 2025; Exner et al., 2025; Lee et al., 2025). On the other hand, public discourse around “AI slop” has accelerated dramatically, with Merriam-Webster naming slop its 2025 Word of the Year and high-profile brand campaigns from Coca-Cola, McDonald’s, and J.Crew facing measurable consumer backlash for AI-generated content perceived as soulless or low-effort (Merriam-Webster, 2025). This paper examines the apparent contradiction against the broader context of what McKinsey (2026) calls the delegation economy, the rapid expansion of consumer and business willingness to delegate decisions and execution to AI agents, and resolves it through a framework we call the Taste Layer.

This research integrates three sources of evidence: a synthesis of the peer-reviewed and industry literature on AI value perception, delegated commerce, and human–AI collaboration in marketing (2023–2026); a qualitative interview study with ten Norwegian Gen Z marketing professionals and business owners (Bekken, 2025); and a longitudinal observational analysis of fifty business accounts using the Native autonomous marketing platform between January and April 2026. The internal cohort data shows that, on average, Native customers post 9.0 times more frequently per month, achieve approximately 5.0 times greater monthly organic reach, and increase total monthly likes by approximately 5.0 times relative to their pre-Native baseline. Ninety-three percent of accounts increased reach, ninety-six percent increased posting frequency, and nine of fifty accounts transitioned from inactive (zero posts per month) to consistently active publishing.

Our central argument has three parts. First, the trust penalty documented in the disclosure literature (Schilke & Reimann, 2025; Koning & Voorveld, 2025) and the performance advantage documented in the field-study literature (Hartmann et al., 2025) are not in conflict: audiences punish content that is perceived as AI-generated, while they reward content that is effectively AI-generated but does not trigger the visual or stylistic heuristics associated with low-effort automation. Second, against the background of expanding delegation to AI agents in commerce (McKinsey, 2026; Deloitte, 2026; Bain & Company, 2025), the marketing function is following the same trajectory: businesses do not want to do marketing, they want the outcomes of marketing, and they will increasingly delegate the work to AI systems that can deliver those outcomes. Third, the competitive moat in this delegation regime is not the underlying generative technology, which has been commoditised, but the taste judgment encoded into the platform itself, the editorial defaults, the brand-extraction quality, and the output filtering that distinguish a platform producing the Hartmann-grade lift from a platform producing slop. We call this combined function the Taste Layer. We ground the framework in the psychology of attention and authorship, integrating the effort heuristic, task-dependent algorithm trust, mind perception, and the recently documented AI-authorship effect, and we show how it is implemented as five operational systems: archetype routing, brand fingerprinting, agentic trend research, acceptance-rate selection, and channel-level performance feedback. Crucially, the Taste Layer is a property of the platform, not a role for the customer. The customer’s job is to operate a coherent business; the platform’s job is to encode taste as product design.

The paper contributes to marketing theory by articulating the Taste Layer as a productisable framework that connects audience value perception research, the persuasion-knowledge model, the emerging literature on delegated and agentic commerce, and the empirical literature on human–AI collaboration in creative work. For practitioners, the paper offers concrete guidance on disclosure strategy, visual aesthetic choices that avoid the AI-tells documented by Exner et al. (2025), and operational design principles that bake taste into the platform layer. We acknowledge methodological limitations including the modest cohort size, the self-selection of customers into Native, the Norwegian and Northern European concentration of the cohort, and the observational rather than experimental nature of the before-and-after design.

Keywords: generative AI, social media marketing, content perception, AI disclosure, AI slop, delegation economy, agentic commerce, audience value perception, taste layer, attention economy, effort heuristic, algorithm aversion, organic reach, platform design

Sammendrag

Den raske integrasjonen av generativ kunstig intelligens i markedsføringsproduksjon har utløst to parallelle og tilsynelatende motstridende utviklinger i det digitale markedsføringslandskapet. På den ene siden viser fagfellevurderte feltstudier at AI-generert visuell markedsføring kan matche og ofte overgå klikkraten til menneskeskapt reklame (Hartmann et al., 2025; Exner et al., 2025; Lee et al., 2025). På den andre siden har den offentlige diskursen rundt «AI slop» akselerert dramatisk, med Merriam-Webster som kåret slop til årets ord 2025, og høyprofilerte merkevarekampanjer fra Coca-Cola, McDonald’s og J.Crew som møtte målbar forbrukermotreaksjon for AI-generert innhold som ble oppfattet som sjeleløst eller halvhjertet (Merriam-Webster, 2025). Denne artikkelen undersøker denne tilsynelatende motsetningen i lys av det McKinsey (2026) kaller delegasjonsøkonomien, den raske utvidelsen av forbrukere og bedrifters vilje til å delegere beslutninger og utførelse til AI-agenter, og løser den gjennom et rammeverk vi kaller smakslaget (the Taste Layer).

Forskningen integrerer tre kilder til evidens: en syntese av den fagfellevurderte litteraturen om verdioppfatninger av AI i markedsføring (2023–2026); en kvalitativ intervjustudie med ti norske markedsførere og bedriftseiere fra generasjon z (Bekken, 2025); og en longitudinell observasjonsanalyse av femti bedriftskontoer som bruker den autonome markedsføringsplattformen Native i perioden januar til april 2026. De interne dataene viser at en gjennomsnittlig Native-kunde publiserer 9,0 ganger oftere per måned, oppnår omtrent 5,0 ganger større månedlig organisk rekkevidde, og øker totalt antall månedlige likerklikk med omtrent 5,0 ganger sammenlignet med før-Native-nivået. Nittitre prosent av kontoene økte rekkevidden, nittiseks prosent økte publiseringsfrekvensen, og ni av femti kontoer gikk fra inaktive (null poster per måned) til konsistent aktive.

Vårt sentrale argument har tre deler. Først er tillitsstraffen som er dokumentert i avsløringslitteraturen (Schilke & Reimann, 2025; Koning & Voorveld, 2025) og ytelsesfordelen som er dokumentert i feltstudielitteraturen (Hartmann et al., 2025) ikke i konflikt: publikum straffer innhold som oppfattes som AI-generert, mens de premierer innhold som faktisk er AI-generert, men som ikke utløser de visuelle eller stilistiske heuristikkene som forbindes med lav-innsats-automatisering. For det andre, mot bakteppet av utvidet delegering til AI-agenter i handelen (McKinsey, 2026; Deloitte, 2026; Bain & Company, 2025), følger markedsføringsfunksjonen samme bane: bedrifter vil ikke gjøre markedsføring, de vil ha resultatene av markedsføring, og de vil i økende grad delegere arbeidet til AI-systemer som kan levere disse resultatene. For det tredje er den konkurransemessige vollgraven i denne delegeringsregimet ikke den underliggende generative teknologien, som har blitt en råvare, men smaksvurderingen som er innebygd i plattformen selv, de redaksjonelle standardinnstillingene, kvaliteten på merkeekstraksjon, og utgangsfiltreringen som skiller en plattform som produserer Hartmann-kvalitet fra en plattform som produserer slop. Vi kaller denne kombinerte funksjonen smakslaget. Vi forankrer rammeverket i psykologien bak oppmerksomhet og opphav, inkludert innsatsheuristikken, oppgaveavhengig algoritmetillit, sinnspersepsjon og den nylig dokumenterte AI-opphavseffekten, og viser hvordan det er implementert som fem operasjonelle systemer: arketyperuting, merkevare-fingeravtrykk, agentisk trendresearch, aksept-rate-seleksjon og kanalnivå-ytelsesfeedback. Avgjørende er at smakslaget er en egenskap ved plattformen, ikke en rolle for kunden. Kundens jobb er å drive en sammenhengende virksomhet; plattformens jobb er å kode smak som produktdesign.

Artikkelen bidrar til markedsføringsteori ved å formulere smakslaget som et produserbart rammeverk som forbinder forskning på publikums verdioppfatninger, persuasjonskunnskapsmodellen, og den empiriske litteraturen om menneske–AI-samarbeid i kreativt arbeid. For praktikere tilbyr artikkelen konkret veiledning om avsløringsstrategi, visuelle estetiske valg som unngår AI-signalene dokumentert av Exner et al. (2025), og operasjonelle designprinsipper som bevarer menneskelig kuratering under forhold med produksjonsskala. Vi anerkjenner metodologiske begrensninger inkludert den beskjedne kohortstørrelsen, selvseleksjonen av kunder inn i Native, den norske og nord-europeiske konsentrasjonen av kohorten, og det observasjonelle snarere enn eksperimentelle ved før-og-etter-designet.

Nøkkelord: generativ AI, sosiale medier-markedsføring, innholdspersepsjon, AI-avsløring, AI slop, delegasjonsøkonomi, agentisk handel, menneske–AI-samarbeid, publikums verdioppfatning, smakslaget, oppmerksomhetsøkonomi, innsatsheuristikk, algoritmeaversjon, organisk rekkevidde, plattformdesign

Preface

This research paper was written by the founding team of Native AS in Oslo, Norway. Native is an AI-native social media automation platform serving small and medium-sized businesses across more than twenty countries. The paper was completed in May 2026, four months after the platform’s public launch and during a period in which the firm grew from a pre-launch product to roughly one thousand paying customers.

The paper builds directly on the master’s thesis of one of the authors, Benjamin Andersen Bekken, completed at the Norwegian University of Science and Technology (NTNU) in June 2025 (Bekken, 2025). That thesis examined how generation Z marketing professionals and business owners in Norway perceive credibility and value in AI-generated versus human-created marketing content. The findings of that work, in particular the conclusion that audiences do not punish AI use per se but instead punish AI use that signals carelessness or absence of human curation, served as the conceptual seed for this paper. The intervening twelve months have produced a substantial body of new empirical evidence, most notably from Hartmann’s research group at TUM and Columbia, Lee and colleagues at Emory and NYU, and Schilke and Reimann at Arizona, which has independently confirmed and significantly sharpened the original qualitative insight.

The motivation for publishing this work under the Native Research banner is twofold. First, the marketing industry is at an unusual moment in which the academic literature, the trade press discourse, and the operational reality of running a marketing tool are giving substantially different answers to the same question: is AI-generated marketing content effective? The peer-reviewed field experiments say yes. The trade press says it depends on whether anyone notices. The operational data from running thousands of accounts says the question is wrongly framed. We wanted to offer a synthesis that takes all three seriously.

Second, Native sits in a category that critics frequently associate with the “dead internet” thesis and the broader concern that AI-generated content is degrading the quality of online discourse. We take these concerns seriously and do not believe the right response is denial or rhetorical defensiveness. The right response is to look honestly at the data, including the data on our own customers, and articulate what conditions distinguish AI-amplified marketing that contributes to attention degradation from AI-amplified marketing that genuinely helps small operators reach the audiences who would benefit from finding them. This paper is our attempt at that articulation.

All three named authors contributed to the conceptual development and the editorial process. Bekken led the literature synthesis and the integration of the original thesis research. Haugå led the operational data extraction and the methodological discussion of measurement and selection effects. Græsberg contributed the technical context on how the Native pipeline operationalises the Taste Layer concept.

We are grateful to the academic researchers whose work we cite throughout this paper. We have endeavoured to represent their findings accurately and to note explicitly where we draw on observational data of our own that should not be confused with peer-reviewed experimental evidence. Any errors of interpretation are our own.

This paper is offered as a research contribution rather than as marketing material. Where we present internal data, we have anonymised customer identities entirely. Where we present conclusions about practice, we have tried to make the conditions and caveats explicit. We invite criticism and disagreement, and we expect the conclusions we reach here will be revised as the field continues to evolve.

Native Research, Oslo, May 2026

1. Introduction

In December 2025, Merriam-Webster announced its Word of the Year. The winner was not a candidate from politics, economics, or popular culture, despite a year filled with all three. The winner was a short, blunt word from the eighteenth century, originally used to describe soft mud and later food waste, now repurposed for the digital era: slop. Merriam-Webster defines the term as “digital content of low quality that is produced usually in quantity by means of artificial intelligence,” citing the flood of “absurd videos, off-kilter advertising images, cheesy propaganda, fake news that looks pretty real, junky AI-written books, workslop reports that waste coworkers’ time… and lots of talking cats” (Merriam-Webster, 2025). The choice of slop over its runners-up, gerrymander, touch grass, performative, tariff, was a cultural statement. In a year of political polarisation and economic disruption, the dictionary editors who track linguistic usage concluded that the dominant anxiety expressed in English-language discourse was about what generative AI was doing to the texture of everyday digital life.

In the same month, the International Journal of Research in Marketing published its most-downloaded paper of 2025. The paper, by Hartmann, Exner, and Domdey at the Technical University of Munich, drew the opposite conclusion. Across three studies escalating from perceptual evaluation through head-to-head creative comparison to a live banner-ad field test running over 173,000 impressions, the authors found that AI-generated marketing imagery can match and frequently exceed human-produced imagery on quality, realism, aesthetics, ad creativity, and, in the field test, click-through rates by up to fifty percent (Hartmann et al., 2025). The authors used the word superhuman in their title, not metaphorically but as a measured empirical claim about how the best generative models performed against commissioned human creative professionals working from identical briefs. A follow-up study from the same research group, using a quasi-experimental “sibling ad” design on a dataset of more than sixteen billion ad impressions and 116 million clicks from a major display ad platform, confirmed the finding at industrial scale (Exner et al., 2025).

The two findings appear to be in direct contradiction. The cultural signal says audiences are turning against AI-generated content with increasing intensity. The empirical signal says AI-generated content is outperforming human-generated content on the metrics that matter most to advertisers. Reconciling these two findings is the project of this paper.

The reconciliation has been hinted at in the literature but has not been articulated as a unified framework. Exner et al. (2025) note that AI-generated ads outperform human-generated ads on click-through rates, but only when those AI-generated ads do not look like AI, a finding that immediately positions the trust penalty as a function of perceived AI authorship rather than actual AI authorship. Schilke and Reimann (2025), in a thirteen-study sequence published in Organizational Behavior and Human Decision Processes, demonstrate that disclosing AI use erodes trust regardless of the underlying quality of the work, and they ground the mechanism in legitimacy theory rather than in performance evaluation. Koning and Voorveld (2025) apply persuasion-knowledge theory to advertising specifically and show that AI disclosure activates the cognitive switch from passive consumption to active evaluation, with predictable consequences for brand attitudes. Bekken (2025), in qualitative interviews with Norwegian Gen Z marketing practitioners, found that audiences themselves articulate this distinction without prompting: AI is not the problem; carelessly executed AI, without the editorial discipline that distinguishes a brand from a bot, is the problem.

A second strand of evidence sets the strategic context within which the marketing-specific findings should be read. McKinsey (2026) documents the rapid emergence of what it calls the delegation economy: consumers and businesses are increasingly willing to delegate decisions and execution to AI agents, with even conservative scenarios projecting that AI agents will mediate three to five trillion dollars of global consumer commerce by 2030. Deloitte (2026) and Bain & Company (2025) report comparable findings on the consumer side, with autonomous shopping assistants moving from experimental to mainstream over the 2025–2026 period. The Stanford AI Index (Stanford HAI, 2025) and Pew Research Center (2025) document the consumer-trust prerequisites for this delegation: trust is fragile, conditional, and asymmetric across categories. The implication for marketing specifically is that the same delegation logic that is reshaping how consumers shop is reshaping how businesses market. Businesses do not want to do marketing; they want the outcomes of marketing, reach, leads, sales, brand awareness, and they will delegate the work to whatever system can reliably deliver those outcomes. The marketing function is being delegated upstream of where it has historically lived in organisational charts.

The unified finding across the AI-perception literature and the delegation-economy literature is that audiences punish signals of AI carelessness rather than the underlying production method, and that businesses are willing to delegate marketing work to systems that deliver outcomes without triggering those signals. The competitive advantage in this regime does not rest on the underlying generative technology, which has been substantially commoditised over 2024–2026. It rests on the taste judgment encoded into the platform itself: the editorial defaults, the brand-extraction quality, the output filtering, and the design choices that determine whether a given piece of generated content reads as brand-distinctive or as generic AI. We call this combined function the Taste Layer. The Taste Layer is a property of the platform, not a role for the customer. The customer’s contribution is to be a coherent business that the platform can extract signal from; the platform’s contribution is to encode taste as product design, and thereby produce output that satisfies both the audience’s demand for non-slop content and the customer’s demand for delegated outcomes.

This paper makes four contributions to the literature and to marketing practice. First, it synthesises the peer-reviewed and industry literature on AI value perception, delegated commerce, and human–AI collaboration in marketing across the rapidly developing period of 2023–2026, integrating field-experimental evidence on AI ad performance, attitudinal evidence on AI disclosure trust effects, qualitative evidence on consumer perception of AI authorship, and the emerging literature on autonomous and agentic delegation. Second, it presents original observational data from a cohort of fifty businesses using the Native autonomous marketing platform between January and April 2026, documenting the operational consequences of AI-amplified marketing when taste is encoded as a platform property rather than as a customer responsibility. Third, it articulates the Taste Layer framework as a productisable concept that links the academic literature to platform design choices. Fourth, it offers concrete strategic guidance for brands, regulators, and platform operators on how to navigate the apparent contradiction between performance gains and trust erosion in AI-mediated marketing.

The central research question of this paper is the following:

RQ: As marketing work is delegated from humans to AI systems, what determines whether the delegated work produces audience approval and business outcomes, or audience backlash and reputational damage?

We address this central question through five subsidiary questions:

Sub-RQ 1. How can the performance findings from field experiments on AI advertising (Hartmann et al., 2025; Exner et al., 2025; Lee et al., 2025) be reconciled with the trust-erosion findings from disclosure experiments (Schilke & Reimann, 2025; Koning & Voorveld, 2025)?

Sub-RQ 2. What visual and stylistic features of marketing content most strongly trigger audience perception of AI authorship, and how does this perception affect downstream engagement?

Sub-RQ 3. What does observational data from a real-world cohort of small and medium-sized businesses using an AI-native marketing platform tell us about the operational consequences of delegating marketing work to a platform on posting frequency, organic reach, and engagement?

Sub-RQ 4. How does the slop discourse documented in trade press and cultural commentary in 2024–2026 map onto the distinct phenomena of engagement farming, brand-led cost-cutting, and population-level content homogenisation, and how do these phenomena relate to the design quality of the platforms producing the content?

Sub-RQ 5. What strategic, operational, and regulatory implications follow from the Taste Layer framework for brands, platforms, and policymakers operating in the delegation economy?

The structure of the paper follows the conventional academic format. Chapter two presents the contextual background, beginning with the acceleration of generative AI in marketing production, then turning to the AI slop discourse and its empirical texture, and concluding with a description of Native AS as the empirical setting for the original data presented later in the paper. Chapter three establishes the theoretical framework, drawing on audience value perception theory, persuasion knowledge theory, the Hartmann field studies, and the emerging literature on human–AI collaboration. The chapter concludes by articulating the Taste Layer framework as the synthesis of these strands. Chapter four describes the methodology used to analyse the Native cohort, with substantial attention to operational definitions, measurement procedures, selection effects, and the limitations of observational before-and-after designs. Chapter five presents the findings from the cohort analysis. Chapter six discusses the implications of these findings for the academic literature, for brand strategy, for regulatory policy, and for the broader trajectory of AI-mediated marketing. Chapter seven concludes with a summary of contributions, a candid statement of limitations, and directions for future research.

Two stylistic conventions deserve note at the outset. First, the language used in this paper has evolved as we have worked on it. Earlier framings centred on terms like AI-amplified marketing and human-led pipelines. The framings that survived to this published version centre on delegation and encoded taste. The shift reflects our increasing conviction that the most consequential AI applications in marketing are not amplifications of human judgment so much as transfers of execution authority from humans to platforms that have taste built into them. The customer’s contribution under this regime is to be a coherent business and to operate the outcome loop (signing up, paying for the service, monitoring results); the platform’s contribution is to do the marketing work. The language reflects this division of labour. Second, we have chosen to publish the paper under the institutional banner of Native Research rather than under a single named author. This reflects the collaborative nature of the work and the fact that the paper draws on operational data and strategic perspectives that none of the named authors could have produced alone. Bekken’s 2025 thesis remains the foundational academic reference for the qualitative perspective; the present paper extends that work substantially with new empirical material and theoretical synthesis.

2. Contextual Background

This chapter establishes the context in which the central question of the paper takes its current shape. We begin with the acceleration of generative AI as a marketing production technology between 2022 and 2026, with attention to the specific developments that have made AI-amplified marketing operationally feasible for small and medium-sized businesses. We then turn to the AI slop discourse, distinguishing three distinct phenomena that are commonly conflated under that label, and documenting the trade press and cultural commentary that has shaped consumer expectations during the period under study. We close the chapter with a description of Native AS as the empirical setting from which the original data presented later in the paper is drawn.

2.1 The Acceleration of Generative AI in Marketing

The integration of generative artificial intelligence into marketing content production has accelerated at a pace that is unusual even by the standards of digital marketing technology adoption. ChatGPT’s release in November 2022 (OpenAI, 2022) is conventionally cited as the inflection point for public awareness of generative AI, but the marketing-specific adoption curve has been driven by a more granular sequence of capability releases across text, image, audio, and video modalities. By the time of writing in May 2026, every major marketing function from copywriting to image generation to short-form video production has at least one generative tool capable of producing output that meets professional baseline quality, frequently at one-hundredth to one-thousandth of the unit cost of human-only production (Hartmann et al., 2025; Kumar et al., 2024).

The economic implications of this cost shift have been substantial. Hartmann et al. (2025) document that across seven state-of-the-art text-to-image models, AI-generated marketing imagery can be produced at a small fraction of the cost of commissioned human creative work while matching or exceeding it on perceived quality, realism, and aesthetics. Their field study of 173,000 real-world ad impressions found that the best-performing AI-generated banner ad achieved a click-through rate approximately fifty percent higher than a professionally commissioned human-made stock photograph for the same advertiser running the same campaign. The follow-up paper by Exner et al. (2025), working with a display ad platform across more than two million ad-day observations and over seven thousand advertisers in nearly fifty product categories, confirmed the finding at scale: ads in which the image was AI-generated outperformed ads with human-generated images on click-through rates, provided the AI-generated images did not visually signal their AI origin.

The conditional clause in the Exner et al. finding is critical and will recur throughout this paper. The performance advantage of AI-generated marketing imagery is not unconditional. It depends on the output not triggering the visual heuristics that audiences use to flag AI authorship. The authors identify several such heuristics empirically. Intense colour saturation signals AI generation to consumers. Hyper-aestheticised compositions signal AI generation. Counter-intuitively, larger human faces and clearer focal subjects, which are stylistic preferences of contemporary generative models, are misattributed by consumers to human creative work. This asymmetry means that AI-generated content can outperform human-generated content while simultaneously being misidentified as human-generated by the audiences engaging with it.

Lee, Todri, Adamopoulos, and Ghose (2025), working independently of the Hartmann group with a different methodology, found a compatible result. In their mixed-methods study of generative AI in advertising, ads created holistically by generative AI consistently outperformed both human-and-genAI-modified ads, increasing click-through rates by up to nineteen percent in field settings. The contrast with genAI-modified ads is theoretically important: when AI is used to tweak existing human work, the performance lift disappears. The lift exists when AI is used to generate from scratch, presumably because the holistic generation process avoids the stylistic inconsistencies that flag AI-modified human work as artificial.

These findings substantially revise the conventional wisdom that AI content is uniformly cheaper but lower quality. The empirical picture is rather that the best AI-generated marketing content, produced by competent operators using current-generation models, is cheaper, faster, and frequently more effective than commissioned human creative work. The constraint on this advantage is not the underlying generation capability but the taste judgment applied at the boundaries of the production pipeline: the encoding of brand and audience signal on the way in, and the editorial filtering of output on the way out. This is consistent with the broader research on human–AI collaboration, which finds that the most effective configurations involve human or human-like judgment at the input and output of an AI execution layer rather than full automation or full manual production (Wilson & Daugherty, 2018; Dell’Acqua et al., 2025).

The strategic consequence for marketing organisations is that the binding constraint on content output has shifted. Five years ago, the binding constraint for most marketing operations was production capacity, the time and money required to produce each individual piece of creative work. The binding constraint today is curation capacity, the time and judgment required to evaluate, refine, and approve creative work that has already been generated. This shift is not yet reflected in most marketing organisational charts, which continue to be structured around production roles (copywriters, designers, videographers) rather than curation roles (creative directors, brand voice editors, taste arbiters). Native’s product design, discussed in section 2.3 and chapter six, can be understood as one operational response to this shift.

It is worth noting what this paper does not claim. We do not claim that generative AI has reached parity with human creative work on all dimensions. There remain marketing tasks for which human creative judgment is clearly superior to current AI capability, brand voice development, long-form strategic copy, complex video narrative, and any work requiring genuine novelty or cultural reference (Mariani et al., 2022; Puntoni et al., 2021). We claim only that on the specific dimension of social media marketing content, where the production task is high-volume, fast-turnaround creative work for established brand voices, the empirical literature increasingly shows that AI amplification of human direction outperforms human production alone on the metrics advertisers actually care about.

Two measurement programmes published after the sources above allow the acceleration to be stated with more precision, and they complicate the most alarmist extrapolations. Ahrefs (2025) analysed 900,000 newly created web pages from April 2025 and found that 74.2 percent contained detectable AI-generated content, but that only 2.5 percent were purely AI-generated with no human involvement; the overwhelming majority of new content is a human-AI blend rather than an unsupervised machine output. Graphite's longitudinal analysis of 55,400 English-language articles sampled from Common Crawl, reported in May 2026, found that the share of new articles that are primarily AI-generated surged to roughly half within two years of ChatGPT's release and has since plateaued near 50 percent for more than a year, while the pages that actually rank at the top of search results remain predominantly human-written or heavily human-edited (Axios, 2026). Both findings matter for this paper's argument. The flood is real, but it has not, so far, followed the exponential path toward the 99 percent scenarios discussed above; and the selection layer, whether algorithmic or human, is already discriminating in favour of curated content. Production has commoditised; selection has not. That asymmetry is the commercial opening the Taste Layer framework formalises.

2.2 The Slop Discourse and Its Empirical Texture

Against this picture of accelerating capability and demonstrated effectiveness, the cultural discourse surrounding AI-generated marketing content has moved in the opposite direction. The term AI slop moved from internet subculture to mainstream cultural vocabulary across 2024 and 2025, with Merriam-Webster naming it Word of the Year for 2025 (Merriam-Webster, 2025) and major news outlets running multiple high-profile features documenting consumer backlash against AI-generated brand content. The Wall Street Journal warned that “AI Slop is Everywhere”; CNET reported that AI slop had “Turned Social Media Into an Antisocial Wasteland”; the Economist had already selected slop as its own word of the year, citing the proliferation of generated short videos in social feeds following the release of OpenAI’s Sora model (Merriam-Webster, 2025; CNN, 2025).

The cultural backlash has produced measurable commercial consequences. Coca-Cola’s 2024 AI-generated Christmas advertisement, intended as a homage to the brand’s 1995 “Holidays Are Coming” spot, was met with sustained criticism for what commenters described as “soulless” and “devoid of any actual creativity” execution (NBC News, 2024). Despite the 2024 backlash, Coca-Cola released a second AI-generated Christmas campaign in 2025, with critics noting the inclusion of animals not typically associated with the winter season (pandas, sloths, seals) as evidence that the AI generation process had not been adequately constrained by editorial judgment (IBTimes, 2025; Contentgrip, 2025). The 2025 ad drew an even sharper response than the 2024 version, with the trade press treating it as a case study in how not to deploy generative AI in brand-defining campaigns.

McDonald’s pulled an AI-generated Christmas advertisement in the Netherlands days after release in 2024 following intense online backlash. J.Crew and Vans collaborated on a sneaker campaign in August 2025 whose AI-generated product images contained distorted hands, warped shadows, and robotic poses that turned what was intended as a launch into a viral meme. In each of these cases, the AI generation itself was technically capable; the failure was in the absence of human curation that would have caught the obvious errors before publication. The pattern across these incidents is what we will later describe as a failure of the curation gate, not of the underlying generative technology.

Table 2.1. Selected high-profile AI marketing controversies, 2024–2025.

It is analytically important to disaggregate the slop discourse into the distinct phenomena it conflates. We distinguish three:

Phenomenon one: Engagement farming. A substantial portion of the AI slop that fills social feeds is produced by accounts that exist solely to harvest algorithmic engagement payouts from platforms such as X and YouTube Shorts. These accounts use AI to generate the imagery and the captions and use bots to post and amplify the content. The economic model is straightforward: trigger the algorithm, capture the rev-share, repeat. The output is recognisable to most observers as obviously low-quality, but the engagement-farming model does not require quality; it requires only enough visual or emotional stimulation to clear the algorithmic threshold. This phenomenon has nothing to do with branded marketing and is unrelated to the operational context in which legitimate businesses use AI to amplify their own content.

Phenomenon two: Brand-led cost-cutting. The high-profile failures documented in Table 2.1 belong to a different category. These are established brands using AI to reduce the cost of campaigns that previously required commissioned creative work, without preserving the human curation that the commissioned process formerly supplied. The failure mode is not the use of AI but the removal of the editorial layer that would have caught the obvious problems before publication. The Coca-Cola 2025 case is particularly instructive because the brand publicly defended its strategy of using “fewer people” in production (Cool Down, 2026), which audiences interpreted not as innovation but as the brand telling them their attention was no longer worth the investment of human craft. The failure is strategic, not technological.

Phenomenon three: Population-level content homogenisation. Even when each individual piece of AI-amplified content is acceptable, the aggregate effect of many brands using overlapping generative models trained on overlapping data is convergence in style, composition, and tone. Doshi et al. (2024) document this empirically in a study of AI-assisted creative writing, finding that while individual AI-assisted writers produced more polished output than unassisted writers, the population-level diversity of output collapsed substantially. The same pattern almost certainly operates in visual marketing content, though the empirical work at scale is still emerging. This is the genuine long-tail risk of widespread AI adoption in marketing, and it is the only one of the three phenomena that cannot be solved purely through better operational practice at the individual brand level. We return to this point in chapter seven.

Figure 2.2. The slop backlash is real and accelerating: AI slop mentions grew from 461,000 in 2024 to 2.4 million in 2025, with 82 percent of sentiment-categorised mentions classified as negative. Source: Meltwater and Brandwatch (2025), as compiled by Native Research.

When critics group Native and similar platforms under the slop label, they are using the rhetoric of phenomenon one (engagement farming) and the visual evidence of phenomenon two (brand cost-cutting failures) to argue against the practice of using AI in marketing at all. This rhetorical move is understandable but analytically incorrect. The three phenomena have different causes, different remedies, and different relationships to the underlying generative technology. A serious response to the slop discourse requires disaggregating them. We attempt that disaggregation throughout the remainder of the paper.

The listening data available at the time of writing indicates that the backlash is a stable regime rather than a passing spike. Brand24's analysis of 228,200 mentions of AI-generated content during May 2026 found 48 percent negative sentiment, with the slop label attached to 35 percent of mentions, and identified AI-generated product imagery as the single most common trigger of negative brand mentions across industries (Brand24, 2026). Survey data moves in the same direction: in a two-wave study of 1,008 United States consumers, the share reporting that heavy AI use would reduce their trust in a brand nearly doubled in a single year, from 20 percent in 2025 to 39 percent in 2026, while 84 percent of consumers want AI-generated content labelled and only 20 percent of organisations report always disclosing it (Search Engine Land, 2026). The gap between labelling demand and labelling practice is itself a finding: the market is operating, almost universally, in the no-disclosure regime whose dynamics chapter four examines in detail.

2.3 The Delegation Economy: Marketing as a Function to Be Delegated

The slop backlash and the field-experimental performance findings discussed above operate against a third backdrop that has emerged sharply over 2025–2026: the rapid expansion of consumer and business willingness to delegate decisions and execution to AI agents. McKinsey (2026) coined the term delegation economy to describe this shift and documented its scope across consumer and enterprise contexts. The relevance for marketing is that the broader pattern of delegation is reshaping the entire question of what businesses want from marketing tools, including the tools studied in this paper.

The consumer-facing evidence is striking. McKinsey (2026) projects, even under moderate scenarios, that AI agents will mediate three to five trillion dollars of global consumer commerce by 2030, with up to one trillion dollars in the United States alone. Deloitte (2026) reports comparable findings: 25 percent of retail sales are expected to flow through AI agents by 2030, with ChatGPT referral traffic already reaching 15–20 percent for some retailers. Bain & Company (2025) documents the financial-rails infrastructure being built to support this delegation, with Visa and Mastercard joining OpenAI and Skyfire to enable agentic payment authorisation. The Linux Foundation’s Agentic AI Foundation, anchored by Anthropic, Block, Google, Microsoft, and OpenAI, was formed in December 2025 specifically to standardise the interoperability and identity protocols needed for autonomous transactions at scale (McKinsey, 2026). Even with consumer-trust concerns documented by the Stanford AI Index (Stanford HAI, 2025) and Pew Research Center (2025), the trajectory is clearly toward expanding rather than contracting delegation.

McKinsey’s (2026) automation-curve framework is particularly useful for thinking about how this delegation unfolds. The framework describes six levels of automation in agentic commerce, ranging from level zero (pre-agentic rules-based subscriptions) through level five (full autonomous delegation). The key insight is that higher levels of automation are not inherently better or more advanced, and the goal is not maximum autonomy but optimal delegation, a level that varies by category, regret risk, and how much human involvement adds value to the experience. Subscribe-and-Save replenishment for household commodities can be delegated almost entirely; luxury purchases tend to retain human involvement. The optimal level of delegation is a function of trust and reversibility, not a fixed value.

Marketing, viewed through this lens, is a function for which most businesses would prefer high delegation. The work itself is not the point; the outcomes are the point. Bekken’s (2025) Norwegian Gen Z marketing-professional interviews surface this preference in qualitative terms: marketers and business owners describe a workflow they would happily delegate if a sufficiently trustworthy system existed. The trade-press research (EMarketer, 2025) shows that marketing teams are beginning to embrace agentic AI for repetitive, high-volume tasks like report generation, competitive monitoring, and content repurposing. The Cro Metrics 2026 marketing predictions (Cro Metrics, 2026) emphasise that 2026 is the learning year in which marketing organisations identify what AI agents do well, with full operational delegation expected to expand sharply through 2027 and beyond. Yahoo DSP, NBCUniversal, Google Ads, and most major ad platforms launched agentic AI features over the 2025–2026 period (EMarketer, 2026), establishing delegation as the default direction of platform development.

The consequence for the present paper is that the marketing function should be analysed within the delegation economy, not within the legacy framework of human marketers using AI as a tool. The legacy framework treats AI as a productivity multiplier for a human role; the delegation framework treats AI as the operator of the role itself, with the human supplying the inputs and consuming the outputs. The small and medium-sized businesses studied in this paper overwhelmingly operate in the second mode. They want sales, leads, and reach. They want to push a button, pay a subscription, and receive the outcome. They have no particular interest in the marketing process; their interest is in its results. This preference is the operational reality the Taste Layer framework has to accommodate, and chapter four develops the framework with this constraint explicit.

One implication of the delegation framing deserves preliminary statement. If marketing is delegated to a platform, then the quality of the marketing outcome is determined by the quality of the platform, not by the quality of the customer’s creative judgment. The customer cannot taste-correct a platform whose taste is bad; they can only switch platforms. This shifts the strategic question from “how does a brand exercise taste through AI tools” to “how does a platform embed taste in its product design.” The McDonald’s, Coca-Cola, and J.Crew failures discussed in section 2.2 are, on this view, not just brand-side failures of curation discipline. They are also tool-side failures: the brands chose tools that did not have adequate taste built in, and applied them in contexts where the absence of taste produced visible damage. The remedy at the brand level is to choose better tools. The remedy at the tool level is to build better products. We return to this distinction throughout the paper.

Behavioural data published during the drafting of this paper indicates that the delegation economy has moved from forecast to measurement. Adobe Analytics, tracking over one trillion visits to United States retail sites, recorded 393 percent year-over-year growth in AI-referred traffic in the first quarter of 2026, following 693 percent growth over the 2025 holiday season, and found that by March 2026 AI-referred visitors converted 42 percent better than visitors from traditional channels, a full reversal from a year earlier when they converted 38 percent worse (Yahoo Finance, 2026). Salesforce estimated that AI agents influenced more than 20 percent of global online retail orders during the 2025 holiday season, and the infrastructure for full delegation arrived in the same window: OpenAI launched native checkout inside ChatGPT in September 2025, and the first litigation over whether autonomous agents may transact on third-party platforms reached a United States federal court in March 2026 (Yahoo Finance, 2026). For the argument of this paper, the direction matters more than the magnitudes: consumers are demonstrably delegating the discovery and comparison stages of commerce to AI systems, and businesses face the same logic on the production side.

2.4 Native AS: The Empirical Setting

The empirical data presented in chapters five and six is drawn from the customer base of Native AS, a Norwegian software company providing AI-native social media automation for small and medium-sized businesses. Native launched publicly in January 2026. The company is headquartered in Oslo and serves customers across more than twenty countries as of the time of writing. The product operates by analysing a customer’s brand, audience, and existing presence, generating content tailored to that brand voice across ten or more publishing channels, and publishing on a cadence determined by the customer’s subscription tier.

The relevance of Native as an empirical setting for this paper is that its product is explicitly designed for the delegation economy regime described in section 2.3. The customer’s active contribution is intentionally minimal: a few minutes of onboarding to authorise platform access to existing accounts and confirm basic brand parameters that the platform has already inferred from the customer’s website and prior presence. From that point forward, the platform operates the marketing pipeline on the customer’s behalf. Signal extraction (brand voice, audience definition, content topics, visual identity) is automated. Generation is automated. Filtering and quality control are automated. Publication is automated. The customer’s ongoing involvement is in the outcome loop: monitoring results, paying the subscription, occasionally rejecting individual outputs that drift off-brand. Most customers do not even do the last of these, and the platform’s design assumes they will not.

The product’s operational results across its customer base provide an empirical test of whether the Taste Layer framework, when operationalised by the platform on the customer’s behalf, produces outcomes consistent with its theoretical predictions, specifically, whether platform-encoded taste can deliver the field-experimental lift documented by Hartmann et al. (2025), Exner et al. (2025), and Lee et al. (2025) in a setting where the customer is not exercising creative judgment themselves.

We make the empirical choice in this paper to anonymise all customer-level data. Customer names, industries, geographies, and any other identifying details have been removed or generalised. This choice reflects both research-ethics considerations and the specific finding from Bekken’s (2025) thesis that customers using AI-amplified marketing tools are sensitive to public attribution of AI involvement. Presenting named customer case studies would have introduced a confound between the operational data we wish to present and the disclosure-effect literature we wish to discuss. We discuss the consequences of this anonymisation choice for the persuasiveness of the empirical findings in chapter five.

The cohort presented in this paper consists of fifty customer accounts that were active on the Native platform between January and April 2026. Cohort selection criteria, the operational definition of “before” and “after”, the measurement procedures for reach, posting frequency, and engagement, and a candid discussion of selection effects are presented in chapter five. The high-level findings, anticipated here for orientation, are that on average across the cohort: posting frequency increased approximately ninefold, monthly organic reach increased approximately fivefold, monthly engagement increased approximately fivefold, ninety-three percent of accounts increased their reach, ninety-six percent increased their posting frequency, and nine of fifty accounts transitioned from inactive to consistently active publication. These findings are observational rather than experimental, and we are careful in chapter six to discuss what they can and cannot tell us about the causal effect of the platform.

3. The Psychology of Attention and AI Authorship

The contextual background presented in chapter two documents two empirical regularities: audiences engage with well-executed AI-generated marketing content at rates matching or exceeding human-made content, and audiences punish content they perceive or are told is AI-generated. Both regularities are behavioural descriptions. Neither is an explanation. This chapter supplies the explanation by examining the psychological mechanisms that govern, first, why people attend to social media content at all, and second, how people evaluate content once questions of authorship and machine involvement enter their awareness. The mechanisms reviewed here are not specific to AI; most were documented decades before generative models existed. Their relevance is that generative AI has changed the economics of content production without changing the psychology of content consumption, and the collision between new production economics and old consumption psychology is precisely what produces the patterns the rest of this paper analyses.

3.1 The Psychological Economics of the Feed

Any account of social media marketing must begin with Simon’s (1971) observation that information consumes attention, so a wealth of information creates a poverty of attention. The social media feed is the purest institutional expression of this scarcity: an effectively infinite supply of content competing for a strictly finite resource. The consequence for marketing is that content does not compete against a quality threshold; it competes against the opportunity cost of the next item in the feed. This framing recasts the value-perception literature reviewed later in this paper: the audience’s implicit question is never “is this good” in the abstract but “is this worth the seconds it costs me relative to what I would see instead.”

Why audiences spend attention on social feeds at all is well described by the uses and gratifications tradition (Katz et al., 1973; Sundar & Limperos, 2013). People come to feeds seeking distinct gratifications: information, social connection, identity expression, entertainment, and habitual mood regulation. Sundar and Limperos (2013) extend the framework to interactive media, noting that the affordances of the medium itself (agency, interactivity, navigability) generate new gratifications beyond those of broadcast media. For marketing content, the implication is that a brand’s post is consumed inside a gratification context the brand does not control. Content that serves the gratification the user came for (a local restaurant’s dish photo serving the browsing-for-dinner gratification) is welcomed; content that interrupts it is screened out or punished.

The compulsive character of feed consumption is explained by the structure of its rewards. Variable-ratio reinforcement schedules, in which rewards arrive unpredictably, produce the most persistent behaviour of any reinforcement structure, a finding established in the operant conditioning literature (Ferster & Skinner, 1957) and deliberately operationalised in product design practice (Eyal, 2014). Each pull of the feed is a lottery ticket: most items are worthless, occasional items are highly rewarding, and the unpredictability itself sustains the checking behaviour. For marketers this has a double edge. The lottery structure guarantees a continuous supply of attention to compete for, but it also means each individual post faces an audience in a rapid sampling mode, allocating evaluation time in fractions of a second before swiping on. Content is therefore judged first on instantly legible surface cues, a fact that becomes central when we turn to AI-detection heuristics in section 3.6.

Three further mechanisms shape what wins inside this sampling regime. Social comparison theory (Festinger, 1954) holds that people evaluate their own opinions and abilities by comparison with others, and feeds are engineered comparison environments; content that offers favourable or aspirational comparison material attracts attention. Parasocial interaction (Horton & Wohl, 1956) describes the one-sided relationships audiences form with media figures, and the social media creator economy is built almost entirely on this mechanism; audiences extend to familiar accounts a relational attention that they do not extend to strangers, which is part of why posting consistency compounds (the account becomes a familiar figure rather than an interruption). Finally, emotion governs both attention and transmission: high-arousal emotional content, whether positive (awe, amusement) or negative (anger, anxiety), is shared substantially more than low-arousal content (Berger & Milkman, 2012), and emotional states themselves transfer through feeds in measurable contagion dynamics (Kramer et al., 2014). The practical summary is that the feed rewards content that is emotionally legible, relationally familiar, and gratification-congruent, all evaluated at sampling speed. Finally, the activation findings reported in chapter six connect to a long-standing concern in the media-literacy literature: Livingstone (2004) and Hoechsmann and Poyntz (2012) describe media literacy as comprising access, analysis, evaluation, and content creation, and the fourth dimension, the capacity to produce media content at all, has been the most unevenly distributed across small businesses, for reasons of cost and skill rather than of motive. The attention economics sketched here is therefore also an economics of participation: lowering the production barrier changes who gets to compete for attention, not merely how effectively the incumbents compete.

3.2 Fluency, Familiarity, and Why Frequency Compounds

Two classic findings explain why the posting-frequency lift documented in chapter six translates into reach and engagement rather than audience fatigue. The first is the mere exposure effect: repeated exposure to a stimulus, even without conscious recognition, increases liking for it (Zajonc, 1968). A brand that appears in a user’s feed twenty times a month is not merely twenty times more available; it is becoming progressively easier to like, because familiarity itself is hedonically positive. The second is processing fluency: stimuli that are easier to process are judged more positively, more truthful, and more trustworthy (Alter & Oppenheimer, 2009). Consistent brand voice, recurring visual identity, and familiar formats all increase fluency, which means a coherent high-frequency presence compounds in a way that an incoherent high-frequency presence does not. The audience is not consciously rewarding consistency; their processing system is rewarding ease.

The fluency mechanism also clarifies a boundary condition for the frequency strategy. Fluency gains accrue to repetition with variation: same voice, same identity, new content. Literal repetition or near-duplicate content tips from fluent into boring, and template-identical content across many brands (the homogenisation phenomenon discussed in section 2.2) erodes the distinctiveness on which brand-level fluency depends. Each individual brand using a generic AI aesthetic enjoys the fluency of that aesthetic only until the aesthetic itself becomes a recognised category, at which point fluency stops attaching to the brand and starts attaching to the category “generic AI content,” with the attributional consequences described in the next sections. This is the psychological restatement of the diversity-collapse problem: fluency is a private good for the first mover and a degraded commons thereafter.

3.3 The Effort Heuristic and Perceived Care

The single most load-bearing psychological mechanism in this paper’s argument is the effort heuristic. Kruger, Wirtz, Van Boven, and Altermatt (2004) demonstrated experimentally that people use perceived effort as a proxy for quality: the identical poem, painting, or piece of armour is judged better, and valued higher, when evaluators believe more time and effort went into producing it, and the reliance on effort cues increases when quality is ambiguous. Social media content is the limiting case of quality ambiguity: evaluation happens in under a second, the evaluator has no expertise in the category, and the content itself rarely admits objective quality measurement. Under these conditions the effort heuristic does not merely influence judgment; it substantially constitutes it.

This mechanism explains the texture of the slop backlash with precision. What audiences describe when they call content “soulless” (the word used both by Coca-Cola’s critics and by Bekken’s 2025 interview participants, independently) is the perception of absent effort: nobody thought about this, nobody checked this, nobody cared whether I saw this. The seven-fingered hands and off-season animals catalogued in Table 2.1 are not aesthetic failures in the first instance; they are effort-cue failures. They communicate that no human spent even the seconds required to notice, and the audience reads that as a statement about how much the brand values their attention. The effort heuristic also explains why the resentment is directed at the brand rather than the technology: the audience’s inference is not “machines make mistakes” but “this brand could not be bothered.” Slop, psychologically, is perceived contempt.

Critically for the Taste Layer framework, the effort heuristic operates on cues of effort, not on effort itself. Kruger et al.’s evaluators never observed the actual production process; they responded to information about it, and in feed contexts the information is carried entirely by the artefact. Content that exhibits the signatures of care (correct details, contextual appropriateness, distinctive voice, compositional restraint) is credited with effort regardless of how it was produced; content that exhibits carelessness signatures is debited regardless of how much labour actually went in. This asymmetry is what makes taste productisable. A platform that reliably ships output carrying high effort-cue density is, in the audience’s perceptual economy, indistinguishable from a brand investing heavy human craft. The effort heuristic is the psychological licence for the paper’s central claim that the curation function can be performed by the platform on the customer’s behalf, because the audience audits the output, never the org chart.

3.4 Algorithm Aversion, Algorithm Appreciation, and Task Dependence

The disclosure literature reviewed in the next chapter documents that audiences punish revealed AI involvement. The psychological literature on algorithmic judgment explains when and why. Dietvorst, Simmons, and Massey (2015) established algorithm aversion: people abandon algorithmic forecasters more readily than human forecasters after seeing them err, even when the algorithm remains demonstrably more accurate. The mechanism is an asymmetric error tolerance; human error is expected and forgiven, machine error is treated as disqualifying. Logg, Minson, and Moore (2019) complicated the picture with algorithm appreciation: in many judgment tasks, lay people actually weight algorithmic advice more heavily than human advice, with aversion concentrated among domain experts and in specific task types. The reconciliation came from Castelo, Bos, and Lehmann (2019), who showed that trust in algorithms is task-dependent: people trust algorithms for tasks perceived as objective (quantifiable, rule-based) and distrust them for tasks perceived as subjective (intuitive, emotional, taste-based). The aversion can be reduced by reframing the task as more objective or by demonstrating algorithmic effectiveness, but the default mapping is robust.

Marketing creative sits squarely in the perceived-subjective category. Audiences understand advertising as an attempt by one party to understand and move another, a task they intuitively classify as requiring human qualities: empathy, cultural fluency, a feel for what lands. The disclosure penalty documented by Schilke and Reimann (2025) and Koning and Voorveld (2025) is therefore not an arbitrary prejudice but the predictable application of the task-dependence rule: an AI doing a subjective task is, in the audience’s implicit taxonomy, an entity operating outside its zone of competence, and its work is discounted accordingly. The same rule predicts the sub-category variation noted by Hartmann et al. (2025): product visualisation reads as relatively objective (render the product accurately) and tolerates known AI involvement, while spokesperson and emotional content reads as maximally subjective and tolerates it least. The task-dependence finding hands the Taste Layer framework an operational rule: route content categories by their perceived subjectivity, and concentrate human-feel signals, and if necessary actual human review, on the subjective end of the spectrum.

3.5 Mind Perception, Anthropomorphism, and the Spokesperson Asymmetry

Why does the audience’s tolerance for AI involvement collapse specifically around emotional and interpersonal content? The mind perception literature supplies the answer. Epley, Waytz, and Cacioppo (2007) describe anthropomorphism as the attribution of humanlike mental states to nonhuman agents, governed by the accessibility of human knowledge, the motivation to explain behaviour, and the desire for social connection. Gray and Wegner (2012) demonstrate that mind perception has two dimensions, agency (the capacity to plan and act) and experience (the capacity to feel), and that machines are granted agency but denied experience. Their key experimental finding is that the uncanny-valley effect is fundamentally about minds rather than appearances: machines become unnerving precisely when they appear to have experience, to feel, because felt experience is the dimension of mind people regard as essentially human.

Marketing content that performs emotion is, in this framework, a claim to experience. A heartfelt founder note, a nostalgic Christmas advertisement, a message of gratitude to customers: these communicate “we feel something toward you.” When the audience believes a machine authored the feeling, the claim is exposed as performed by an entity that, in the audience’s ontology, cannot feel, and the result is the specific discomfort Gray and Wegner identify, experienced in marketing contexts as the “soulless” reaction. The Coca-Cola Christmas failures are best understood through this lens: the brand chose the most experience-claiming content category in its calendar (warmth, nostalgia, togetherness) as the venue for visible automation, maximising the mismatch between the emotional claim and the perceived author. The same campaign executed as product-focused content would, on this analysis, have drawn a fraction of the backlash.

3.6 Authenticity, Moral Disgust, and the AI-Authorship Effect

The mechanisms above converge in the newest experimental work on AI-generated marketing communications specifically. The psychology of authenticity has long shown that people value objects and works partly through their causal history rather than their observable properties: an original artwork is valued far above a perfect duplicate because evaluators treat the creator’s actual performance as part of the object (Newman & Bloom, 2012). Jago (2019) extends this to organisational conduct, finding that work attributed to algorithms is perceived as less authentic than identical work attributed to humans, because authenticity attributions track the presence of a genuine expressing self behind the act.

Kirk and Givi (2025) provide the most direct evidence in the marketing communication context, across seven preregistered experiments. When consumers believe an emotional marketing communication was written by AI rather than a human, they perceive it as less authentic, experience measurably higher moral disgust, and reduce both positive word of mouth and loyalty; the authors term this the AI-authorship effect, serially mediated by perceived authenticity and moral disgust. The moderation pattern is at least as important as the main effect. The penalty is attenuated for factual rather than emotional messages, attenuated when the AI only edits rather than authors the communication, attenuated when the message is signed by the AI itself rather than passed off under a human signature, and actually reverses when consumers believe the communication was reused or copied rather than originally written, because in that comparison the AI version is no less authentic than the alternative. The emotional involvement of disgust matters here: disgust is an avoidance emotion, which is why the behavioural consequences extend beyond attitude (lower trust) into action (reduced advocacy, switching).

The Kirk and Givi moderation pattern is a near-exact experimental confirmation of the mind-perception and task-dependence logic developed above, and it sharpens the Taste Layer framework in three ways. First, it identifies emotional register, not content category in general, as the variable that gates the authorship penalty: factual, informational, and product-focused content is comparatively safe terrain for visible automation, while first-person emotional expression is the danger zone. Second, the deception component matters independently: the penalty is worst when machine work wears a human signature, which means the perceived attempt to pass is itself part of the offence. Third, the reversal under the reuse comparison shows that the baseline against which AI content is judged is not an idealised artisanal human but the realistic alternative; where the realistic alternative is recycled boilerplate, AI authorship carries no authenticity deficit at all. For the small-business cohort studied in this paper, whose realistic alternative was sporadic posting or silence rather than commissioned craft, this third finding is arguably the most consequential in the entire psychological literature.

3.7 Detection Psychology: Cues, Confidence, and the Moving Frontier

The final mechanism concerns the audience’s actual, as opposed to believed, capacity to detect AI authorship. The detection literature shows a wide and widening gap between the two. Nightingale and Farid (2022) found that AI-synthesised faces had become indistinguishable from real faces for ordinary observers, and were in fact rated as more trustworthy than real faces, a result the authors attribute to synthetic faces approximating averaged, prototypical features that observers process fluently. Industry survey work finds roughly half of consumers believe they can identify AI-generated content (Innovation Visual, 2025), a confidence the experimental accuracy data does not support, particularly for current-generation models. Detection, in practice, is not perception of artificiality; it is heuristic inference from a small set of learned surface cues, the saturation, gloss, and suspicious perfection documented by Exner et al. (2025).

Three properties of this heuristic detection regime matter for the framework. First, because detection runs on cues rather than on ground truth, both error types are common: cue-free AI content passes as human (the condition under which the Hartmann-grade performance lift is earned), and cue-bearing human content is misread as AI (photographers and illustrators publicly accused of using generators became a recurring 2025 phenomenon). The operative variable throughout this paper is therefore perceived AI likelihood, a continuous psychological quantity, not binary authorship. Second, the cue set is socially learned and therefore drifts: as audiences absorb each generation of tells through memes and commentary, yesterday’s invisible output becomes today’s flagged output, which means anti-detection aesthetics are a maintained capability rather than a one-time design choice. Third, the asymmetry of consequences identified in sections 3.3 through 3.6 means the cost of a false “AI” classification is high (effort-cue collapse, authenticity penalty, disgust) while the reward for a confident “human” classification is modest. Rational platform design therefore optimises not for the average perception but for minimising the tail probability of triggering the cue set at all. This is the psychological specification of the output-filtering stage of the Taste Layer.

3.8 Summary: Psychological Microfoundations of the Taste Layer

Table 3.1 summarises the mechanisms reviewed in this chapter and maps each to its implication for AI-amplified marketing and to the stage of the Taste Layer pipeline (developed in the next chapter) where it binds. The unifying observation is that every mechanism operates on perceivable properties of the output and its presentation, never on the hidden facts of production. Attention is allocated by surface legibility, value is inferred from effort cues, trust is gated by perceived task fit and perceived authorship, and authenticity is judged against the realistic alternative. Production method enters audience psychology only by way of the signals it leaves behind. That is the deep reason the apparent contradiction this paper set out to resolve is resolvable at all: the performance literature and the backlash literature are observing the same psychology applied to different signal profiles.

Table 3.1. Psychological mechanisms, implications, and Taste Layer pipeline stages.

With the psychological foundations in place, the next chapter assembles them, together with the marketing-specific perception, disclosure, field-performance, collaboration, and delegation literatures, into the Taste Layer framework proper.

4. Theoretical Framework

This chapter develops the theoretical framework used to analyse the empirical material presented in chapter six and to ground the strategic implications discussed in chapter seven. The framework draws on four established research streams in marketing and communication theory: audience value perception (Zeithaml, 1988; Saha, 2024), persuasion-knowledge theory (Friestad & Wright, 1994; Koning & Voorveld, 2025), source credibility and legitimacy (Metzger & Flanagin, 2013; Schilke & Reimann, 2025), and human–AI collaboration in creative work (Wilson & Daugherty, 2018; Dell’Acqua et al., 2025; Doshi et al., 2024). The chapter concludes with a synthesis we call the Taste Layer framework, which articulates the conditions under which AI-amplified marketing produces audience approval rather than backlash.

4.1 Audience Value Perception and the Quality–Source Tradeoff

The foundational theoretical question in this paper is how audiences assign value to marketing content they encounter in social media feeds. Zeithaml’s (1988) classic definition treats perceived value as “the consumer’s overall assessment of the utility of a product based on perceptions of what is received and what is given.” The framework distinguishes four operational conceptions of value: value as low price, value as whatever the consumer wants, value as quality relative to price, and value as the tradeoff between benefits received and costs given. The fourth conception is the most analytically useful for digital marketing content, where the relevant currency is attention rather than money, and the relevant cost is the opportunity cost of the audience’s time.

Saha’s (2024) exploratory work on AI content perception extends this framework to AI-generated marketing material. Saha finds that audience value judgments are multidimensional, combining informational utility, personal relevance, emotional connection, and authentic representation. The relevant cost dimensions include time, attention, perceived privacy intrusion, and, importantly for AI content, the cost of feeling deceived if AI authorship is perceived after the fact. Click Consult’s (2024) survey work documents the trade structure quantitatively: while eighty-one percent of respondents expressed privacy concerns about AI in marketing, forty percent simultaneously believed AI would make marketing more creative in 2025. The same respondents are not contradicting themselves; they are weighing benefit and cost dimensions that point in different directions.

Bekken’s (2025) qualitative work with Norwegian Gen Z marketing practitioners surfaces the same tradeoff structure inductively. Across ten in-depth interviews, participants converged on two dominant value criteria when evaluating marketing content: content quality and brand reputation. Production method was secondary. AI involvement was not viewed as inherently devaluing; it became devaluing only when associated with poor execution, generic templating, or absence of editorial care. One participant in Bekken’s study captured the position with unusual clarity:

“If I don’t see it as generated, then it’s good generation because there’s been a human behind it too. But you see, or I mean at least, that poor prompting and poor use of AI makes the content you see soulless because it doesn’t feel like there’s a person who has thought it out. It feels like it’s just following goals and just following best practices for everything. And we humans aren’t like that.”, Belinda, Head of Operations, in Bekken (2025)

This participant’s position is consistent with the broader value-perception literature: AI is evaluated as a tool, and the tool is evaluated on the output it produces. The relevant cost dimension is not “AI was used” but “the brand did not care enough to ensure the output reflected its values.” This is the operational meaning of the abstract concept audience-perceived effort. Audiences read the level of effort invested in content as a signal of how much the brand values their attention, and they reciprocate accordingly. AI-generated content that reflects evident editorial discipline reads as effort; AI-generated content that visibly does not reads as contempt. Critically, the audience cannot distinguish between editorial discipline supplied by the brand’s in-house team and editorial discipline supplied by the platform on the brand’s behalf. What audiences detect is the result of editorial discipline, not its source.

The implication for the present paper is that the relevant analytical unit is not “AI versus human” but “well-curated versus poorly-curated.” The empirical findings from Hartmann et al. (2025), Exner et al. (2025), and the Native cohort data presented later in this paper are all consistent with this reframing. The same is true of the negative findings on AI disclosure, to which we now turn.

4.2 Persuasion Knowledge and the Disclosure Paradox

The persuasion knowledge model (Friestad & Wright, 1994) provides the theoretical mechanism that explains why AI disclosure erodes trust even when the underlying content is high quality. The model proposes that audiences develop, through repeated exposure to persuasion attempts, a cognitive system for recognising and evaluating persuasive intent. Activation of this system triggers a shift from passive content consumption to active evaluation of the communicator’s motives and tactics. The shift has predictable consequences for attitudes toward the message: messages whose persuasive intent is salient are evaluated more critically and produce smaller attitudinal changes than messages whose persuasive intent is non-salient.

Koning and Voorveld (2025) apply this framework directly to AI-generated advertising in a between-subjects experiment with 304 participants. Their finding is that AI disclosures increased conceptual AI knowledge and attitudinal persuasion knowledge, with the consequence of decreased trust toward both the advertisement and the organisation behind it. The mechanism is persuasion-knowledge activation: disclosure flips a cognitive switch from “consuming content” to “evaluating a persuasion attempt,” and the evaluation is unfavourable. Importantly, the effect is on the labelling rather than on the underlying production method. The content quality has not changed; the audience’s relationship to the content has.

Schilke and Reimann (2025) provide the most comprehensive empirical treatment to date of the AI disclosure trust penalty. Their thirteen-study sequence in Organizational Behavior and Human Decision Processes spans diverse contexts, communications, analytics, artistry, supervision, education, and finds that actors who disclose their AI usage are trusted less than those who do not, with the effect holding across both individual actors (supervisors, subordinates, professors, analysts, creatives) and organisational actors (investment funds). The mechanism they propose, drawing on micro-institutional theory, is legitimacy: disclosure undermines perceived legitimacy, and legitimacy mediates trust. The trust penalty is attenuated, but not eliminated, among evaluators with favourable technology attitudes and beliefs in AI accuracy.

The Schilke and Reimann findings have several important features that bear on this paper’s argument. First, the penalty is robust across framings of disclosure: whether AI use is disclosed generally (“AI was used”) or specifically (“AI was used but the human revised the output”), the trust penalty persists. Second, the penalty is robust whether disclosure is voluntary or mandatory: legally compelled disclosure produces the same trust erosion as self-disclosure. Third, and most striking, individuals condemn others for using AI even when they themselves use AI in similar tasks. The asymmetry suggests that the legitimacy mechanism operates at the level of social norm enforcement rather than at the level of individual cost-benefit calculation.

The Nuremberg Institute for Market Decisions (NIM, 2025) ran what is now the most-cited industry-scale experiment on this question. Using a representative sample, NIM presented two halves of the sample with identical product advertising, one half labelled as a photograph and the other half labelled as AI-generated. The result confirmed the experimental literature: simply knowing that a piece of content was crafted by an algorithm rather than by a human creative caused respondents to trust it less and engage with it less enthusiastically. Technically polished AI content faced what the NIM researchers termed the “trust penalty”: a bias whereby consumers react warily when they sense a message was created by a machine.

The disclosure paradox is now empirically well-documented. Audiences punish disclosed AI use even when the underlying content is identical to undisclosed content. The strategic and regulatory implications are non-trivial and have not been resolved in the policy literature. The European Union’s Digital Services Act and emerging AI Act provisions on transparency assume that disclosure protects consumers; the empirical record suggests that disclosure damages the disclosing party with consequences that may also reduce the supply of content that consumers actually value (Hermann & Puntoni, 2025). We return to this tension in chapter seven.

For the present chapter, the analytically critical point is this: the disclosure penalty operates on perceived AI authorship, not on actual AI authorship. The same physical content produces different audience responses depending on whether the audience has been told it was AI-generated. This is the foundation on which the next subsection builds.

Since the studies above were conducted, the disclosure literature has grown large enough to support systematic synthesis. Baryshkov (2026), following PRISMA protocols, reviewed 35 empirical studies of consumer trust responses to AI-generated marketing content published between 2020 and 2026 and reached conclusions that align closely with the framework developed in this chapter and the psychological mechanisms reviewed in chapter three: AI disclosure activates persuasion knowledge and erodes trust-related outcomes across diverse marketing contexts, the effects are real but neither universal nor uniform, and perceived authenticity emerges as the primary mediating mechanism, with moral disgust operating as a parallel affective path. Additional experimental work has extended the disclosure penalty to services advertising, where Grigsby, Michelsen and Zamudio (2025) found across three experiments that AI disclosures lowered both trust toward the provider and attitudes toward the advertisement. The convergence is worth stating plainly: an independent systematic review of the field arrives at the same mediation structure, authenticity first, disgust alongside, context moderating, that this paper's framework predicts.

4.3 The Hartmann Findings and the ‘Looks Like AI’ Constraint

The work of Jochen Hartmann’s research group at the Technical University of Munich, in collaboration with Columbia Business School and Harvard Business School colleagues, provides the most rigorous empirical evidence to date on the field performance of AI-generated marketing content. The body of work spans three principal studies: Hartmann, Exner, and Domdey (2025) on the perceptual and creative quality of AI marketing imagery; Exner, Hartmann, Netzer, Zhang, and Ding (2025) on the field performance of AI-generated display ads at industrial scale; and the follow-up identification of the specific visual features that signal AI authorship to consumers.

Hartmann et al. (2025), published in the International Journal of Research in Marketing and recently identified as the journal’s most-downloaded paper of 2025, proceeds through three studies that escalate in ecological validity. Study one is perceptual quality: 254,400 human evaluations across 10,320 synthetic images generated by seven state-of-the-art text-to-image models (DALL-E 3, Midjourney v6, Firefly 2, Imagen 2, Imagine, Stable Diffusion XL Turbo, Realistic Vision), benchmarked against 2,400 real human-made marketing images. The finding: AI-generated marketing imagery can surpass human-made images in quality, realism, and aesthetics. Study two is a creative head-to-head: identical creative briefings were given to commissioned human freelancers and to the AI models. Result: the best synthetic images excelled on ad creativity, ad attitudes, and prompt following. Study three is a field test: a banner ad campaign with an online education provider running over 173,000 impressions compared AI-generated images against professional human-made stock photography. The headline result was that AI-generated banner ads achieved up to a fifty percent higher click-through rate than a human-made image. The authors’ concluding claim is that the paradigm shift brought about by generative AI enables marketing content production at superhuman effectiveness levels, with substantial implications for advertising economics.

Exner et al. (2025) extended the finding to industrial scale. Working with a major display ad platform, the authors leveraged a quasi-experimental “sibling ad” design, comparing 4,633 AI-generated and human-made visuals launched by the same advertisers within identical campaign settings at the same time. Same brand, same campaign, same timing, same audience: only the production method differed. The dataset comprised more than two million ad-day observations across over seven thousand advertisers in nearly fifty product categories, encompassing more than sixteen billion ad impressions and 116 million clicks. The headline finding was that AI-generated ads outperformed human-generated ads on click-through rates, but only if the AI-generated images did not look like AI. The conditional is the entire contribution of the paper.

The authors identify several visual features that signal AI authorship to consumers. Intense colour saturation signals AI generation. Hyper-aestheticised compositions signal AI generation. Counter-intuitively, larger human faces signal human rather than AI authorship, even though current generative models produce these features more frequently than commissioned human creative work. The mismatch between what AI actually does and what audiences expect AI to do creates an exploitable asymmetry: AI-generated content that violates the AI-tells can be misattributed to human creative work, and this misattribution preserves the engagement that disclosure would destroy.

Figure 2.1. Engagement crossover: AI content outperforms human benchmarks when not perceived as AI-generated, but trust and engagement collapse as perceived AI likelihood approaches one hundred percent. Adapted from Hartmann, Exner, and Domdey (2025) and Exner et al. (2025).

Lee, Todri, Adamopoulos, and Ghose (2025) provide independent confirmation from a different research group, different platform, and mixed-methods design. Their key finding: AI-created ads consistently outperformed both human-created and AI-modified ads, increasing click-through rates by up to nineteen percent in field settings. AI-modified ads, by contrast, showed no significant improvement over human-created benchmarks. The interesting theoretical contribution is the workflow asymmetry: AI delivers greater value when used for holistic ad creation rather than for modification of human work. This may be because holistic generation avoids the stylistic inconsistencies that flag hybrid AI-modified work as artificial, or because the comparative test sets up an unfavourable baseline for modification studies. Either way, the finding is consistent with the broader Hartmann result: AI-generated content can outperform human-generated content, conditional on the AI process being done well.

The strategic implication of this body of work is that the constraint on AI marketing performance is not whether to use AI but how to use AI. The same underlying technology produces opposite outcomes depending on whether the output trips the AI-detection heuristics. The constraint is operational and editorial rather than technological. This is the second pillar on which the Taste Layer framework rests.

4.4 Human–AI Collaboration and Diversity Collapse

The third theoretical pillar is the literature on human–AI collaboration in creative and knowledge work. Wilson and Daugherty’s (2018) early framework of “collaborative intelligence” proposed that the most effective configurations of AI in work settings were those in which humans and AI specialised in their respective strengths: humans on judgment, strategy, and contextual sensitivity; AI on pattern recognition, generation at scale, and consistency. The intervening years have substantially refined the picture with empirical data.

Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon, and Lakhani (2025) report results from a large field experiment at Boston Consulting Group involving 758 consultants assigned to AI-augmented and control conditions across a battery of consulting tasks. Their headline finding is a productivity “jagged frontier”: AI assistance dramatically improved performance on tasks within the AI’s current capability frontier (roughly twenty-five percent faster, with output quality forty percent higher than the control group) but degraded performance on tasks beyond the frontier, where consultants who relied on AI produced worse output than those who did not. The implication is that the value of AI in creative work depends on accurate judgment about when to rely on AI and when not to, a judgment that itself requires human expertise.

Doshi, Hauser, and Mollick (2024) provide the empirical evidence for the most worrying finding in this literature, which we will refer to as “diversity collapse.” In a large pre-registered experiment with creative writers, the authors found that AI assistance improved the quality of individual stories produced by less-creative writers, with the effect concentrated at the lower end of the creative ability distribution. However, the population-level diversity of stories collapsed. Writers using the same AI tool produced stories with greater stylistic and thematic similarity to each other than writers working without AI assistance. The mechanism is not difficult to understand: generative models are trained on large datasets and produce output that reflects the central tendencies of their training distributions. Users who accept the model’s suggestions converge on those central tendencies. Users who deviate from the model’s suggestions, the “taste layer” in our terminology, preserve diversity, but only to the extent that they actively exercise the judgment to deviate.

The diversity collapse finding has direct implications for the AI slop discourse. Phenomenon three (population-level content homogenisation) is not an accusation; it is an empirically documented property of large-scale AI adoption in creative work. The aggregate effect of many brands using overlapping generative models trained on overlapping data is convergence in style, composition, and tone. The defence against this convergence is not refusal to use AI, a strategy that is increasingly economically untenable for small and medium-sized businesses, but disciplined exercise of the curation function that preserves brand-distinctive output despite the convergent gravitational pull of the underlying models.

A related finding from the field-experimental literature is reported by Liu, Sun, and colleagues (2025) in a multi-million-impression field experiment on the X platform comparing human-human teams against human-AI teams. The headline finding was that human-AI collaboration improved click-through rates on text quality and view-through duration, while human-human teams remained superior on image quality measures, with the consequence that human-AI teams were significantly more task-oriented, producing twenty-five percent more task-oriented messages and eighteen percent fewer interpersonal messages. The same study, however, found that human-AI collaboration produced more homogeneous outputs at the population level, replicating the Doshi et al. (2024) finding in a different domain.

Across these studies, a consistent picture emerges. Human–AI collaboration is most effective when human judgment is concentrated at the points in the workflow where it adds the greatest value: at the input boundary, where the brand’s identity, audience, and intent are encoded; and at the output boundary, where work is accepted, refined, or rejected. AI is most valuable when it handles the high-volume execution work between these two boundaries. The configurations that fail are those that remove human judgment from both boundaries (producing slop) or that fail to leverage AI between them (leaving the productivity gains on the table).

A clarification of practical importance follows. The human contribution at the input boundary does not need to take the form of a written creative brief authored by the customer. In the small and medium-sized business context that is the empirical setting of this paper, customers do not typically have the time, the expertise, or the inclination to write briefs. Most SMB customers want to push a button and receive marketing that produces sales; they are paying precisely to avoid the marketing-direction labour that a written brief would require. What the literature on human–AI collaboration actually requires is not a brief but a sufficient signal about what the business is, who it serves, and what it should sound like. That signal can be supplied passively, through a customer’s existing website, their existing social presence, the products they sell, and the language they already use to describe themselves, and extracted by the platform on their behalf. The platform does the brief work; the customer supplies the latent signal by existing as a coherent business. This distinction matters for how we describe the framework, and we make it explicit in the next section.

4.5 Delegation, Autonomy, and the Limits of Customer Effort

The fifth theoretical pillar, less developed in the marketing literature than the four above but increasingly important for understanding the operational context in which marketing tools are now deployed, is the emerging body of work on delegation to AI agents. This work spans economics, organisational behaviour, and consumer research, and has developed rapidly during 2025–2026 as autonomous AI systems have moved from experimental to mainstream deployment.

The conceptual core of the delegation literature is the observation that humans differ systematically in the tasks they wish to perform versus the tasks they wish to delegate. The earliest formal treatment in modern AI contexts is Russell’s (2019) work on the alignment problem in agent design, which frames delegation as a question of whose preferences the agent serves and how those preferences are inferred or specified. Empirically, McKinsey (2026) documents what it calls an automation curve across consumer commerce, ranging from level zero (pre-agentic subscriptions and replenishment) to level five (full autonomous delegation across discretionary purchases). The key empirical claim is that the optimal level of delegation is not maximal; it varies by category, reversibility, and how much value human involvement adds to the experience. Some tasks people want to delegate fully; some they want to retain; the boundary between the two is a function of trust, regret risk, and how much the act itself is intrinsically valued.

For marketing, the implications are direct. Marketing is rarely intrinsically valued by the businesses that pay for it. Bekken (2025) documents this attitude inductively in qualitative interviews; industry surveys (EMarketer, 2025; Cro Metrics, 2026) document it quantitatively at scale. Marketers and business owners describe marketing as work they perform to obtain outcomes they care about, sales, leads, reach, not as work they would perform for its own sake if the outcomes could be obtained otherwise. This is the precondition for delegation: the function is instrumental rather than expressive. Where the function is purely instrumental, delegation expansion is limited only by the trust prerequisites Bain & Company (2025), Stanford HAI (2025), and Pew Research Center (2025) document as fragile but improving. Where the function is intrinsically valued, for example, in artisan craft markets where the act of making is the point, delegation is bounded by the desire to remain involved.

This distinction has substantial implications for the design of AI tools in marketing contexts. A tool designed on the assumption that the customer wants to be a creative director is solving the wrong problem for most SMB customers. A tool designed on the assumption that the customer wants the marketing function delegated to a trustworthy system is solving the right problem, provided the tool actually performs the function well enough that the customer’s outcomes do not suffer. The trustworthiness question becomes central. McKinsey’s (2026) automation-curve work emphasises that delegation is reversible: a customer who experiences a delegation failure (a McDonald’s-tier output, a J.Crew-tier visual error, a brand voice violation) will withdraw their delegation and either revert to lower-automation tools or change providers. Sustaining delegation therefore requires sustaining the conditions under which the delegated outcomes are reliably good.

The autonomy-versus-control tension is also documented in the broader human–AI interaction literature. EY’s (2025) consumer trust research found that 66 percent of respondents say human oversight remains essential for AI use, even as the same respondents reported delegating significant tasks to AI agents. The apparent contradiction resolves under the same framework McKinsey applies to commerce: respondents want oversight to be available without wanting to exercise it on routine tasks. They want the option of intervention, not the obligation. This is a design constraint for delegation-economy products: they must be auditable, reversible, and overridable, even if customers in practice rarely audit, reverse, or override.

Applied to the Taste Layer framework, the delegation literature contributes the following: taste does not have to be a customer responsibility in order to be effectively exercised. It can be a platform property, provided the platform is designed to maintain the trust prerequisites that sustain delegation. This is the reframing we apply to the synthesis in the next section.

4.6 Synthesis: The Taste Layer Framework

The five research streams reviewed above converge on a framework that we call the Taste Layer. The central claim of the framework is that, in AI-mediated marketing, the competitive moat does not lie in the underlying generative technology (which has been commoditised), nor in the human-creative-direction work performed by the customer (which most customers do not want to perform and will increasingly delegate away), but in the taste encoded into the platform itself, the editorial defaults, brand-extraction quality, and output filtering that distinguish a platform producing Hartmann-grade lift from a platform producing slop. Taste, on this framing, is a product property, not a role.

This is a deliberate departure from earlier framings of human–AI collaboration in marketing, which typically assumed that the human in the loop was the customer exercising creative judgment. The earlier framings were not wrong as descriptions of how marketing worked in large brand organisations with internal creative teams. They were wrong as descriptions of how marketing works in the small and medium-sized business segment, which has grown to dominate the customer base of AI-native marketing platforms and is the empirical focus of this paper. In the SMB segment, the customer does not have a creative team. The customer does not want to be a creative team. The customer wants to push a button and receive outcomes. The Taste Layer framework accepts this as the operational starting point and asks: under what conditions can a platform deliver high-quality marketing outcomes when the customer is in the delegation regime rather than the creative-direction regime?

The framework decomposes the AI marketing pipeline into four functional stages, with explicit attention to where taste judgment is exercised and by whom:

Stage one: Signal extraction. The brand’s identity, audience, voice, and intent must be encoded into a representation the generation layer can work from. In the large-brand regime, this signal is provided by the customer in the form of brand guidelines, creative briefs, and content strategy documents. In the SMB delegation regime, the platform extracts the signal from the customer’s existing presence, website copy, prior social posts, product descriptions, customer reviews, the language the business actually uses. The signal is what the business is, not what the customer has to articulate. The quality of this extraction is the first major design lever distinguishing platforms with taste from platforms without.

Stage two: AI execution. Given a signal, the generation layer produces content at scale across modalities. Hartmann et al. (2025), Exner et al. (2025), and Lee et al. (2025) document that this stage can produce output at or exceeding professional human creative quality when the signal is good and the generation is done well. The technological capability at this stage has been commoditised across major foundation models; the differentiation lies in how the generation is conditioned on the signal, not in which model is used. This is the second major design lever.

Stage three: Output filtering. Generated content is evaluated against quality standards, brand-consistency criteria, factual constraints, and contextual appropriateness checks, with non-conforming output rejected, refined, or replaced. In the SMB delegation regime, this filtering is performed by the platform rather than by the customer, since asking the customer to perform it would defeat the purpose of delegation. Filtering can be implemented through algorithmic checks, internal editorial review, AI-on-AI evaluation, or hybrid combinations. The McDonald’s, Coca-Cola, and J.Crew failures discussed in section 2.2 are filter-stage failures: the underlying generation was technically capable, but the filtering was inadequate or absent, and obvious problems reached audiences. This is the third major design lever and, in our judgment, the most important.

Stage four: Audience reach and engagement. Published content is distributed through algorithmic platforms and produces measurable outcomes. When stages one through three are executed well, this stage produces the performance lift documented in the field-experimental literature and reflected in our cohort data. When any of stages one through three fails, particularly stage three, this stage produces the slop-discourse backlash.

Figure 4.1. The Taste Layer framework: a four-stage pipeline in which taste is encoded as platform property at stages one and three. The customer’s contribution is to be a coherent business and to operate the outcome loop. The platform’s contribution is to extract signal accurately at stage one, to generate competently at stage two, and to filter aggressively at stage three.

Several implications of the framework deserve emphasis.

The customer is in the outcome loop, not the production loop. The Taste Layer framework is a delegation framework. The customer’s contribution is to be a coherent business that the platform can extract signal from, to authorise the platform to publish on their behalf, and to monitor outcomes. The platform’s contribution is to perform the marketing work. This division of labour is not a compromise relative to a hypothetical ideal in which the customer exercises creative direction; it is the operational regime most SMB customers actively prefer, and the framework treats this preference as a feature to be designed for rather than a problem to be solved.

Taste becomes a product property. In legacy frameworks of human–AI collaboration, taste was a property of the human in the loop. In the Taste Layer framework, taste is a property of the platform: it is encoded in how the signal is extracted, how the generation is conditioned, what is rejected at the filter, and what editorial defaults the platform applies when no explicit customer instruction is available. This is a design discipline, not a labour input. It is also a source of competitive advantage that durably distinguishes platforms in a category where the underlying generative technology is commoditised. The investment differentiator for AI marketing tools, on this framework, is taste engineering: the editorial design, the brand-extraction algorithms, the rejection criteria, and the quality control that determine what the platform will and will not ship on the customer’s behalf.

The competitive advantage is durable in a way the generation advantage is not. Foundation models will continue to converge. The marginal generation capability of one foundation model relative to another is shrinking. What does not converge is the editorial discipline, brand-extraction sophistication, and filtering rigour that one platform brings versus another. Two platforms using the same underlying foundation models can produce dramatically different outputs depending on how they have designed their signal-extraction layer, conditioning approach, and filter. This is where the Taste Layer creates durable differentiation. It is also where the slop-versus-quality distinction is settled.

The framework provides a natural account of the disclosure paradox. Disclosure of AI involvement activates persuasion knowledge regardless of the underlying quality of the work (Schilke & Reimann, 2025; Koning & Voorveld, 2025). From the audience perspective, the disclosure does not communicate “this brand is using a high-taste AI platform.” It communicates “this brand is using AI,” and the audience’s prior associations with the label dominate. The strategic implication is that platforms and the brands they serve should make their taste visible through the output itself rather than through disclosure language: through the consistency, distinctiveness, and contextual appropriateness of what they publish. The brand voice is the disclosure. The taste is the disclosure.

The framework offers an account of the diversity collapse problem that locates the remedy at the platform layer. If many brands operating in similar markets use overlapping AI models with similar default outputs, population-level convergence is the predictable result (Doshi et al., 2024). On legacy framings, the remedy was for brands to invest in distinctive briefs. On the Taste Layer framework, the remedy is for platforms to invest in distinctive signal extraction and filtering, such that two customers with genuinely different businesses produce genuinely different outputs even when the underlying generation model is shared. The customer-level remedy still operates for brands with the resources to exercise it; the platform-level remedy is what scales to the SMB segment.

The framework as stated is functional: it says what must happen at each stage, not how. The next section closes that gap by mapping the pipeline onto the five operational frameworks through which the platform implements it in production.

4.7 From Mechanism to Product: Five Operational Frameworks

The framework developed in the preceding section is articulated at the level of functional stages: signal extraction, execution, filtering, outcome. Stated at that level, “taste as a product property” risks reading as metaphor. This section descends one level of abstraction and describes the five operational frameworks through which the platform studied in this paper implements the pipeline in production. The purpose is not product description for its own sake. It is to demonstrate that every psychological mechanism reviewed in chapter three has a corresponding engineering control, that the taste layer is not a slogan about human oversight but a set of named, monitored systems, each of which can fail independently and each of which is measured. Because several of these systems constitute proprietary engineering, we describe each at the level of its function and its controlling signals rather than its internal algorithms; the level of description is sufficient to evaluate the theoretical claims, which is the standard this paper holds itself to. The five frameworks are: brand archetype (what kind of business this is), tone of voice and brand fingerprinting (what this business sounds and looks like), trends (what the world around the business is talking about), taste (what a human curator accepts), and social performance (what the market of the customer’s own channel rewards).

4.7.1 Brand archetype: classification before generation

The first framework operates at the moment of signal extraction. When a customer’s data is consolidated at onboarding, a dedicated classification pass types the business along structural dimensions, language, brand type, writing style, and archetype, and the downstream generation strategy is routed by that classification. An e-commerce brand is routed toward product-led visual strategies; an expertise business toward knowledge-led strategies; a personality-led brand toward founder-forward formats. The classification is not cosmetic: it determines which content categories the system will attempt at all, and in which emotional registers.

The psychological warrant comes directly from the task-dependence literature reviewed in section 3.4 and the mind-perception findings in section 3.5. Audience tolerance for visible automation varies systematically by content category: product visualisation reads as a comparatively objective task and tolerates machine involvement, while first-person emotional expression is the maximum-risk register (Castelo et al., 2019; Gray & Wegner, 2012; Kirk & Givi, 2025). A routing rule needs a routing table, and the archetype is that table. Knowing what kind of business a brand is tells the system which categories are safe terrain, which require the most conservative treatment, and which gratification the brand’s audience is plausibly in the feed to obtain. The archetype framework is, in effect, the task-dependence rule compiled into the platform.

4.7.2 Tone of voice and brand fingerprinting: identity as a binding constraint

The second framework deepens the extracted signal from a category to an identity. The brand fingerprint is extracted from the customer’s existing presence (website, prior posts, product imagery, visual assets) and stored as a structured object covering visual identity, colour application, compositional complexity, preferred production styles, aesthetic keywords, typographic and logo integrity, alongside textual voice dimensions such as tone, structure, and writing style. Two implementation properties matter more than the field list. First, the fingerprint is applied structurally rather than advisorily: it enters the generation pipelines as binding parameters, a single source of truth that downstream layers cannot quietly override, rather than as a textual hint a generative model is free to ignore. Second, the fingerprint is treated as a living hypothesis about the brand: a degraded extraction is retried rather than persisted, and repeated “off-brand” declines from the customer trigger a review of the fingerprint itself, the system questions its understanding of the brand, not merely the individual outputs.

Psychologically, the fingerprint is the mechanism that makes posting frequency compound instead of fatigue. Section 3.2 established that fluency gains accrue to repetition with variation, same voice, same identity, new content, and that an incoherent high-frequency presence forfeits the compounding entirely. A structured, enforced identity is what guarantees the “same voice” half of that formula at machine scale. The fingerprint is also the platform-level answer to diversity collapse (Doshi et al., 2024): conditioning generation on a brand-specific representation is precisely the discipline that moves output away from the mean of the model’s training distribution, so that two customers with genuinely different businesses produce genuinely different content even when the underlying foundation model is shared. Distinctiveness, on this design, is not a customer virtue; it is enforced at the conditioning layer.

4.7.3 Trends: agentic search as input replenishment

The third framework keeps the extracted signal current. Generation is grounded in continuously refreshed external material: industry research, web and news search relevant to the specific business, and periodic re-analysis of the customer’s own properties, performed agentically by the platform rather than requested from the customer. The architectural insight behind the framework is a diagnosis: when the novelty yield of generation declines for a brand despite variation pressure, the binding constraint has migrated from the generation layer to the input layer. The brand’s existing knowledge has been mined out. The correct response is targeted acquisition, ask the customer for a recent win, research the industry for a timely angle, revisit the business’s own properties for new material, not recycling of the existing corpus.

Three mechanisms from chapter three converge here. Attention economics (section 3.1) implies that content competes against the opportunity cost of the next item in the feed, and the next item is current; topicality is part of gratification congruence at sampling speed. The fluency boundary condition (section 3.2) states that repetition with variation requires a supply of new variation, which is an input problem before it is a generation problem. And Kirk and Givi’s (2025) reuse baseline cuts in the platform’s favour: the realistic alternative to fresh, researched content for the small-business segment is recycled boilerplate or silence, the comparison condition under which AI involvement carries no authenticity deficit at all. The trends framework exists so that the platform’s output is never competing against that baseline from the wrong side.

4.7.4 Taste: acceptance rate as a measured property

The fourth framework is the deepest, and it is where the paper’s central claim, that taste can be a product property, becomes literal measurement. Every generated suggestion is presented to a human, one at a time, for acceptance, decline, or edit, and every interaction is durably recorded together with a structured description of the content it concerned. Acceptance statistics are then aggregated per content pattern over rolling windows of days to weeks, scored with deliberately conservative statistics that refuse to reward small-sample flukes. On those scores, the system applies continuous selection pressure: content patterns whose acceptance collapses are automatically retired; new candidate patterns are derived from high-performing ones under explicit variation pressure; and, decisively for the argument of this paper, every system-proposed change passes a human review gate before going live. The pattern is system proposes, human approves, instituted at the platform level in exact parallel to the customer’s per-post gate. The framework is completed by guardrails that encode epistemic humility as policy: an automated quality gate screens newly derived patterns before any customer is exposed to them; content classification operates under an abstain rule (low confidence means do not record and do not learn, never guess); and a fail-open invariant guarantees that no failure in a learning component can ever block or degrade generation itself.

This is the effort heuristic (section 3.3), productised. Kruger et al.’s (2004) finding was that audiences value content through perceivable cues of care, audited from the artefact alone; the acceptance rate is the platform turning that same audit on its own output, continuously, with a human curator as the instrument. Every decline is a recorded effort-cue failure; every retired content pattern is a pattern the curation gate has measured as falling below the care threshold audiences enforce. The framework is equally the engineering form of the detection-psychology conclusion in section 3.7: if rational design optimises the tail probability of triggering the audience’s AI-cue set, then a recurring selection cycle that retires cue-bearing patterns and derives variation from cue-free ones is that optimisation, run as infrastructure. Acceptance rate is to the taste layer what click-through rate was to the field experiments reviewed in section 4.3: the operational variable through which an otherwise abstract quality, care, craft, taste, becomes a number that can fall, trigger an alarm, and be acted on.

4.7.5 Social performance: the channel as final arbiter

The fifth framework closes the loop at the pipeline’s fourth stage. Curator acceptance is a proxy; the criterion is the market response of the customer’s own channel. Post-publish analytics, reach, impressions, and engagement, retrieved per platform over time, feed back into the same learning substrate as acceptance events, so that what a brand’s actual audience rewards can reweight what the system generates next. The framework also maintains independent health signals that preference data alone cannot supply: a per-brand variety target, and concentration measures over recent output that alarm on thematic narrowing before the audience does.

The justification for treating these health signals as first-class is one of the more striking patterns to emerge from the platform’s own production data. An internal measurement across a large production corpus of brands (Native internal data, Q2 2026) found that, for a given brand, the theme the system generated most was systematically accepted at a substantially lower rate than starved minority themes, in the focal case, the dominant theme was accepted roughly half as often as the under-supplied ones. Generation share was negatively associated with acceptance. This is diversity collapse and fluency degradation observed live, inside a single brand’s pipeline: the over-mined theme had crossed the boundary described in section 3.2, from fluent to boring, while the under-supplied themes retained their novelty value. A learning system that optimised acceptance alone would have reproduced the collapse one brand at a time, by feeding each brand more of what it had already accepted until the audience tired of it. The social-performance framework, with its variety target, is the institutional correction: the platform deliberately maintains deviation from its own learned preferences, because the psychology of section 3.2 predicts, and the production data confirms, that pure preference-following self-defeats. As with the cohort data in chapter six, this measurement is observational, drawn from a single measurement window, and reported in aggregate; we present it as a pattern consistent with the theory rather than as a causal estimate.

4.7.6 Two loops, one taste layer

The five frameworks resolve into a clean architecture. Archetype and the brand fingerprint constitute signal extraction: who this brand is, compiled into routing decisions and binding constraints. Trends constitute input replenishment: what the brand has to say next. The remaining two are complementary control loops. The acceptance-rate framework is a preference loop, it answers “what does this brand and its curator want?” and guards against off-brand and low-care output, operationalising fluency and the effort heuristic. The social-performance framework, together with the variety measures, is a health loop, it answers “is generation healthy regardless of preference?” and guards against the failure mode that pure preference learning would itself create. Table 4.2 maps each framework to its pipeline stage, the psychological mechanism it operationalises, and its implementation signature. The unifying observation parallels the one that closed chapter three: every mechanism in audience psychology operates on perceivable properties of the output, and every framework above exists to control a perceivable property of the output. The taste layer is the sum of those controls, which is why it can be engineered, measured, and, as the next chapter tests, evaluated against field outcomes.

Table 4.2. The five operational frameworks mapped to pipeline stages, psychological mechanisms, and implementation signatures.

With the framework articulated at both the functional and the operational level, we now turn to the methodology used to test its implications against the Native cohort data.

5. Methodology

This chapter describes the methodology used to analyse the Native customer cohort introduced in section 2.3. We are explicit throughout the chapter about the observational, non-experimental nature of the data and about the consequences this has for the inferences that can be drawn. The chapter is organised in six sections: research design, data sources and cohort definition, the operational definition of the before-and-after comparison, measurement procedures, selection effects and limitations, and ethics and anonymisation.

5.1 Research Design

The research design used in chapters five and six is observational longitudinal. For each customer in the cohort, we observe performance metrics (posting frequency, reach, engagement) over a period before the customer began using Native and over a period after the customer began using Native. We compare the two periods. This design is sometimes called a pre-post or interrupted-time-series design (Shadish, Cook, & Campbell, 2002), and it is a standard design for observational analysis of platform interventions when randomised assignment is not feasible.

The design is observational rather than experimental in two important senses. First, customers were not randomly assigned to Native; they self-selected into the platform by purchasing a subscription. Second, the “before” and “after” periods are not counterfactual: we observe the actual post-adoption trajectory of each account, but we do not observe what would have happened to the same account had it not adopted Native. Both features of the design introduce limitations that we discuss in section 5.5. The findings we report should be read as describing the experience of accounts that used Native during the study period, not as causal estimates of what Native does to a randomly selected business.

We have considered alternative designs that would have stronger causal identification, including matched-pair quasi-experimental designs (matching Native users to similar non-users by industry, size, and pre-Native performance) and interrupted-time-series designs with control series. Both are valuable and represent natural extensions of the work presented here. We chose to publish the simpler observational analysis in this paper because the directness of the comparison serves the paper’s broader analytical purpose: showing what the lift looks like in operational practice. We are explicit about the inferential limitations of this choice in section 5.5 and again in chapter six.

5.2 Data Sources and Cohort Definition

The cohort consists of fifty customer accounts using the Native platform during the observation window from January 2026 through April 2026. Cohort selection proceeded in three stages.

First, we identified all customer accounts active on Native as of 1 April 2026 with a paid subscription tenure of at least sixty days. This minimum-tenure criterion was imposed to ensure that each account had sufficient post-adoption data to support a meaningful before-and-after comparison. As of 1 April 2026, Native had approximately one thousand active paying customers; the sixty-day-tenure subset comprised approximately three hundred accounts.

Second, we restricted the candidate pool to accounts for which we had access to platform-side analytics covering both the pre-adoption and post-adoption periods. This restriction reflects a real measurement constraint: for accounts whose social media handles were inactive or non-existent before adopting Native, the “before” period contains no meaningful data, and we treat these separately as new-activation cases (see section 6.4). For accounts whose handles existed and were measurable, we drew on the social media platforms’ native analytics APIs to retrieve reach, posting frequency, and engagement data for both periods.

Third, we selected the fifty accounts with the longest and most complete data series, prioritising accounts that had been on the platform for at least three full months and whose pre-adoption period contained at least sixty days of measurable activity (or that had no pre-adoption activity at all, allowing classification as new-activation cases). This is a convenience sample rather than a random sample, and the consequences of the selection procedure are discussed in section 5.5.

The cohort spans seven industry segments. Table 5.1 summarises the distribution. We have anonymised industry labels at a slightly higher level of abstraction than the underlying customer industries to reduce identifiability risk while preserving the cohort’s diversity.

Table 5.1. Cohort composition by industry segment (n = 50).

Geographic composition is concentrated in Norway and Northern Europe. Forty-three of the fifty accounts (eighty-six percent) operate primarily in Norway, with the remainder distributed across Sweden, Denmark, the United Kingdom, the Netherlands, and Germany. This geographic concentration reflects Native’s commercial origin in the Norwegian market and the early-stage cohort’s composition; it is not representative of the broader global business population and should not be treated as such. We discuss the consequences for external validity in section 5.5.

5.3 Defining Before and After: The Index Date Convention

For each customer, we define an index date as the date on which the customer’s first content piece was published through the Native platform. The pre-Native period is defined as the ninety calendar days immediately preceding the index date. The post-Native period is defined as the period from the index date through 30 April 2026, with a minimum length of sixty days enforced by the cohort selection criteria. The index date convention is administratively clean and avoids the complication of partial-month effects that would arise from a calendar-month-based definition.

There is an obvious methodological objection to this convention: business activity is not uniformly distributed in time, and the ninety days immediately preceding a customer’s decision to purchase a marketing tool may be systematically different from the ninety days immediately following. A business that is investing in marketing capability is typically a business with momentum, and momentum is correlated with the pre-period measurements. We address this concern in two ways. First, we test sensitivity to the choice of pre-period length by re-computing the headline lift metrics using 60-day, 90-day, and 120-day pre-windows; the topline lift magnitudes are robust within approximately ten percent of the headline numbers reported in chapter six. Second, we explicitly include and report the new-activation cases (n = 9) in which the pre-period contained zero activity. These cases would, on a strict before-and-after design, produce undefined lift ratios; we treat them separately as a distinct phenomenon (see section 6.4).

For accounts with non-zero pre-period activity but limited data (fewer than ten posts in the ninety-day pre-window), we report results both with and without these accounts included to verify that the topline lift is not driven by low-baseline accounts. The robustness checks are reported in section 6.5.

5.4 Measurement of Reach, Posting Frequency, and Engagement

The three primary metrics analysed in this paper are posting frequency, organic reach, and engagement. Each is operationally defined as follows.

Posting frequency is defined as the count of distinct content pieces published on a customer’s primary social media account during the relevant period, divided by the number of weeks in the period and multiplied by a standard four-week month to produce a per-month figure. The metric is computed separately for each platform and then summed across platforms to produce a total-posts-per-month figure for each customer. We use the standard four-week month rather than calendar months to allow direct comparison across windows of different lengths.

Organic reach is defined as the number of unique accounts that the platform’s analytics reports as having been served the customer’s content during the relevant period, summed across all of the customer’s active platforms, and averaged to a per-month figure using the same four-week-month convention. Reach is operationalised separately from impressions to avoid double-counting users who saw a given piece of content multiple times. We rely on the platforms’ own reach reporting and acknowledge that platforms compute reach with methods that are not fully transparent and that may change over time. The consequence of this opacity is that the reach metric is best interpreted as a within-account comparison (the same platform measures the same account before and after) rather than as a cross-account or cross-platform comparison.

Engagement is defined as the sum of likes, comments, shares, saves, and other affirmative interactions on the customer’s content during the relevant period, averaged to a per-month figure. For ease of communication, we typically report engagement as “monthly likes” since likes are the dominant engagement category by volume in all platforms in the cohort; the lift magnitudes are similar (within a few percentage points) when engagement is computed with the full multi-category definition.

Table 5.2. Operational definitions of key metrics.

5.5 Selection Effects and Limitations

Honesty requires us to be explicit about the limitations of this design. We discuss four principal concerns: self-selection of customers into Native, survivor bias in the cohort, confounding with other marketing investments, and external-validity limitations from the geographic and tenure profile of the cohort.

Self-selection. Customers who purchase Native are not a random sample of small and medium-sized businesses. They are businesses that have decided their current marketing capability is insufficient and that have selected an AI-amplification approach to address the gap. The empirical consequence is that the pre-period measurements likely understate the trajectory the business would have achieved without Native, because the decision to adopt Native is itself correlated with a perception that the trajectory needed to change. Conversely, the post-period measurements are consistent with a business that has decided to invest in marketing capability and is doing so. The lift we measure is therefore the lift achievable by businesses who decide to invest in AI-amplified marketing, not the lift that would be achieved by businesses randomly assigned to do so. This is a real limitation and we do not paper over it. The lift is operationally interesting because it describes what self-selected adopters experience, but it should not be read as a causal estimate of the Native treatment effect on a counterfactual non-adopter.

Survivor bias. Our cohort consists of accounts that were still active on Native at the cohort selection point (1 April 2026). Customers who churned before this point are excluded. If churn is correlated with poor performance, which it almost certainly is in any subscription-based product, then the cohort is biased toward better-performing customers. We have considered including churned customers and re-computing the headline metrics on the full population of customers with at least sixty days of tenure regardless of whether they remained subscribed. We do not present that analysis here because the data infrastructure to support it at the time of writing is incomplete: we have aggregate churn rates but not the platform-side analytics for churned accounts that would be needed to compute their lift on a comparable basis. We commit to publishing this fuller analysis in a follow-up paper. In the meantime, the figures reported here should be read as representative of retained customers, not of the full adopting population.

Confounding. Customers who adopt Native typically also do other things to invest in their business, including hiring marketing staff, increasing other marketing spend, redesigning their websites, and launching new products. The lift we observe over the post-Native period combines the effect of Native with the effect of all these confounding investments. We do not attempt to isolate the Native-specific component of the lift. The honest description of what we measure is the trajectory of accounts that adopted Native, not the causal effect of Native alone.

External validity. As noted in section 5.2, the cohort is concentrated in Norway and Northern Europe. The audience composition of the platforms (Meta, LinkedIn, TikTok, etc.) varies by geography, and the baseline performance of small-business accounts varies by market. The lift magnitudes we report are likely to vary across geographies, and we make no claim that the topline lift figures generalise to small-business accounts in geographies with very different platform dynamics (for example, China or India, where the dominant platforms differ substantially from the European mix). We treat the findings as evidence about what AI-amplified marketing can produce for self-selected adopters in Northern European small-business contexts, and we are explicit about the geographic boundary of the inference.

What the design can tell us. The observational design cannot tell us what Native does causally to a randomly assigned business. It can tell us what self-selected adopters’ trajectories look like before and after adoption, and whether those trajectories are consistent with the theoretical predictions of the Taste Layer framework articulated in chapter four. The framework predicts that AI amplification operating through a pipeline that preserves signal extraction at the input and curation at the output should produce substantial gains in posting frequency, reach, and engagement, because the binding constraint on these metrics for the cohort businesses prior to adoption was production capacity, not audience demand. The findings reported in chapter six are consistent with this prediction. They do not prove the prediction; they are consistent with it. This is a more modest empirical claim than is sometimes made for internal-cohort studies in marketing tool advocacy, and we believe the modesty is warranted by the design.

5.6 Ethics and Anonymisation

All customer-level data presented in this paper has been anonymised. Customer names, account handles, geographic locations (beyond country-level aggregation), and industry sub-categories have been removed. Numerical metrics have been rounded or expressed as ratios in ways that prevent re-identification of individual accounts. In Figure 6.1 (Selected Anonymised Lift Profiles), we present ten cases by index letter (A through J) rather than by customer identifier; the index letters are randomly assigned and do not correspond to alphabetical ordering of customer names.

The decision to anonymise was made for two reasons. First, the literature reviewed in chapter four demonstrates that public disclosure of AI involvement in marketing content can erode brand trust regardless of the underlying quality of the work (Schilke & Reimann, 2025; Koning & Voorveld, 2025). Named customer case studies would have created a disclosure event for those customers and potentially harmed their brand standing in ways unrelated to the merits of their actual content. Second, customers using AI-amplified marketing tools have privacy-like interests in not being publicly identified as such, parallel to the interests of customers using any other operational software. We have honoured those interests rather than seeking case-study consent.

The empirical consequence of this choice is that the cohort findings are less rhetorically powerful than they would be with named customer case studies and quotes. We accept this tradeoff in service of the broader analytical argument the paper is making about the disclosure paradox. A research paper that argued against prominent disclosure while itself making prominent attribution to customers would have been internally inconsistent.

All Native customers were informed at sign-up that aggregate, anonymised data about their account performance may be used for product development and research purposes. The data presented in this paper falls within the scope of that disclosure. No customer-identifying data has been shared outside the Native organisation. The analysis underlying this paper was conducted internally by the named authors using internal platform tooling. We have not engaged external auditors for this paper and we acknowledge that internal analysis is a weaker epistemic position than independently audited analysis would be. Future work will benefit from external replication or audit.

6. Findings

This chapter presents the findings from the cohort analysis described in chapter five. We organise the findings around the three primary metrics defined in section 5.4: posting frequency, organic reach, and engagement. We then present the new-activation cases separately because they represent a qualitatively distinct phenomenon. We close with selected anonymised lift profiles to give the reader a sense of the within-cohort variance, and with an honest section on what the data cannot tell us. Throughout, we remind the reader that the design is observational and the magnitudes should be read as descriptions of adopter trajectories rather than as causal treatment effects.

Table 6.1. Topline performance metrics, Native 50-customer cohort (January–April 2026).

The headline numbers are presented as approximate cohort averages with the share of accounts improving on each metric. Lift ratios are computed at the account level and then averaged across the cohort. New-activation cases (accounts with zero pre-period activity) are excluded from the ratio averages because the lift ratio is undefined when the denominator is zero; they are reported separately as the row in the table indicating that nine of fifty accounts transitioned from zero baseline activity to consistent posting in the post-Native period.

6.1 Posting Frequency: A Nine-Fold Lift

The largest single lift in the cohort is in posting frequency. On average across the cohort, accounts publish approximately nine times more content per month in the post-Native period than they did in the ninety-day pre-Native period. Ninety-six percent of accounts increased their posting frequency; the two accounts (four percent) that did not were accounts that had been at very high pre-Native posting cadences for unusual reasons (an agency-managed period that ended around the index date, in one case; a paused product launch in the other) and that returned to a more typical small-business cadence post-adoption.

The magnitude of the posting-frequency lift is the easiest of the three primary metrics to interpret theoretically. The cohort consists overwhelmingly of small and medium-sized businesses whose marketing function is staffed at the level of zero to one part-time person. The binding constraint on posting frequency for these businesses is production capacity, the time required to conceive, produce, and publish each individual content piece. Native’s product collapses this cost: once the brand voice and content strategy are established at onboarding, the marginal cost of each additional piece is the few seconds of human attention required at the curation gate plus the platform-side compute cost. The empirically observed nine-fold lift is consistent with the theoretical prediction that removing the production constraint should produce a large frequency response.

The strategic significance of the posting-frequency lift is that it changes the regime in which the account operates on the algorithmic platforms. Social media platform algorithms reward consistent posting because consistent posting is associated with active account stewardship and audience-relevant publishing (Macready, 2024). Accounts that post sporadically are treated less favourably by algorithmic distribution than accounts that post consistently, with the consequence that the same content piece produces less reach when published into an algorithmically-suppressed account than when published into an algorithmically-rewarded account. The posting frequency lift is therefore not just a quantity story; it is also a quality-of-distribution story, and it sets up the reach lift documented in the next section.

6.2 Reach: The Five-Fold Average and the 93 Percent Floor

Monthly organic reach across the cohort increased by an average factor of approximately five times. Ninety-three percent of accounts saw an increase in reach. The seven percent of accounts (three accounts) whose reach did not increase did so for traceable reasons: in one case, the account had had a viral pre-period due to a single piece of content that did not represent the account’s typical reach, and the post-period was lower than the viral pre-period but higher than the pre-period excluding the viral piece. In the other two cases, the accounts had reduced their publication on platforms with declining reach (in particular, Facebook) in favour of platforms with higher engagement (TikTok and Instagram), and the reduction in Facebook reach was not fully offset by the increase elsewhere. We consider these cases honestly representative of the variance within the cohort and do not exclude them from the topline.

The reach lift is the metric that most directly speaks to the audience-side effects of AI-amplified marketing. Reach is the quantity that determines how many potential customers see a brand’s content; it is the upstream of every downstream marketing outcome. A five-fold lift in reach for a small business operating on a tight marketing budget is operationally significant; it is the difference between a marketing presence that is plausibly competitive in a local market and a marketing presence that is functionally invisible. The fact that ninety-three percent of accounts improved on this metric, essentially the entire cohort, indicates that the lift is not driven by a few high-performing outliers but is a property of the cohort as a whole.

It is worth noting how the reach lift relates to the posting frequency lift. The two are not the same metric. A nine-fold increase in posts does not mechanically produce a five-fold increase in reach, because the per-post reach typically declines as posting frequency increases (the platform distributes a limited per-account daily reach budget across more posts, and the marginal post receives fewer impressions than the inframarginal post). The empirically observed ratio, reach scales by approximately the square root of post frequency in the cohort, with substantial variance, is consistent with the algorithmic dynamics that platform observers have described informally (Borkakoty, 2025). The substantive interpretation is that the cohort is using AI amplification to publish much more content, that the per-post reach declines as expected but does not collapse, and that the net effect is substantial aggregate reach growth.

We caution against over-interpreting reach as a measure of marketing success in its own right. Reach is a leading indicator of business outcomes (sales, leads, brand awareness) rather than a direct measure. A five-fold lift in reach that produces a one-percent lift in sales is operationally different from a five-fold lift in reach that produces a fifty-percent lift in sales, and the cohort data we present here does not distinguish the two cases. Native does not have systematic access to customers’ sales data, and we make no claim about downstream conversion lift. The reach lift is reported as a reach lift, not as a profit lift.

6.3 Engagement: Likes Scale With Reach

Monthly engagement across the cohort increased by an average factor of approximately five times, closely tracking the reach lift. The proportionality between reach lift and engagement lift indicates that the per-impression engagement rate is roughly preserved across the pre and post periods. This is theoretically interesting: if the AI-amplified content were systematically less engaging than the pre-Native human-only content, we would expect per-impression engagement to decline, and the engagement lift would lag the reach lift. The cohort data does not show this pattern. Engagement scales with reach in approximately the proportion expected if the content quality is at parity with the pre-Native baseline.

This finding is consistent with the Hartmann et al. (2025) and Exner et al. (2025) results discussed in chapter four: AI-generated content, when produced through a well-curated pipeline, performs at least at parity with human-only content on engagement-related metrics. In the field experimental literature, the AI content frequently outperforms; in our observational data, we cannot distinguish parity from a small advantage because the noise in the per-impression engagement rate is comparable to the magnitude of the effect we would expect. We report engagement-scaling parity as a robust finding and decline to claim that the AI-amplified content is engagement-superior to the pre-Native baseline. The honest claim is that the AI amplification does not appear to have degraded engagement quality, and the more aggressive claim of engagement superiority would require either a larger cohort or a more carefully controlled design.

The composition of engagement across categories (likes, comments, shares, saves) is broadly stable between the pre and post periods, with one noteworthy exception. Comments per impression tend to be modestly lower in the post-period than in the pre-period for some accounts, particularly in the professional services segment. We interpret this carefully. One possible interpretation is that AI-amplified content is less conversational and therefore prompts fewer comment responses; another is that the post-period contains a higher proportion of broadcast-style content (announcements, product features) and a lower proportion of community-engagement-style content (questions, polls, replies). We have not done the content-coding work to distinguish these interpretations rigorously. The decline in comment rate is small (typically five to fifteen percent of the pre-period baseline) and is more than offset by the larger reach base. But it is worth flagging as a candidate for the diversity-collapse phenomenon discussed in chapter four: if AI amplification tends to push content toward broadcast modes and away from conversational modes, the long-run consequences for audience relationship may differ from the short-run consequences for reach and likes.

6.4 Activation: Nine Accounts From Zero to Active

Nine of the fifty accounts in the cohort had zero or near-zero social media activity in the ninety-day pre-Native period and transitioned to consistent activity in the post-Native period. We treat these as a separate phenomenon from the lift cases because the lift ratio is mathematically undefined when the denominator is zero, and because the qualitative experience these businesses report is different from the experience of businesses that scaled up an existing marketing presence.

Across the new-activation cases, the post-Native posting frequencies range from approximately eight to thirty posts per month and the monthly reach figures range from approximately ten thousand to roughly thirty thousand monthly accounts. These figures are substantial for accounts that had previously been functionally absent from social media. In post-adoption interviews conducted as part of routine customer success contact (and reported here in aggregate rather than as named cases), the new-activation cohort described their pre-Native state in consistent terms: they had not previously published on social media because they did not have the time, did not feel confident about content production, or had attempted social media presence in the past and abandoned it after initial efforts did not produce engagement. The post-Native experience was described as the first time these businesses had had a sustainable social media operation.

The new-activation cases are theoretically significant because they speak directly to the access dimension of media literacy discussed in chapter three. Livingstone (2004) and Hoechsmann and Poyntz (2012) describe media literacy as encompassing four dimensions: access, analysis, evaluation, and content creation. The fourth dimension, the capacity to produce media content, has historically been the most unevenly distributed across the population of small businesses. The cost and skill barriers to producing professional social media content have meant that many small businesses simply do not participate in social media as producers, even when they could benefit from participating. AI amplification, operationalised through a product like Native, substantially lowers the access barrier on the fourth dimension. The empirical question of whether this access expansion is socially valuable or socially neutral is open and depends on what the new participants publish. The new-activation cases in our cohort published business-relevant content for their respective markets, and the audiences in those markets appear to have valued the content as evidenced by the engagement figures. We do not generalise from this to a claim that all access expansion is valuable, but we do note that the cases in our data are consistent with a positive interpretation.

6.5 Distribution Profiles: Selected Anonymised Lift Curves

To give the reader a sense of the variance underlying the cohort averages, we present ten anonymised lift profiles in Figure 6.1 and Table 6.2. The profiles were selected from the top decile of cohort gainers and are not representative of the full cohort distribution; they are presented to illustrate the upper tail rather than to characterise the centre. The table reports pre-period and post-period monthly reach for each profile, with the lift ratio computed in the standard way. Two of the profiles (D and G) are flagged as new-activation cases where the pre-period reach is at or near zero.

Figure 6.1. Selected anonymised lift profiles: monthly reach before and after Native (Native 50-customer cohort, Q1 2026). Profile D is a new-activation case (pre-period reach of zero). Profile G is effectively a new-activation case (pre-period reach of approximately five). Profile labels A through J are randomly assigned and do not correspond to alphabetical order of customer names.

Table 6.2. Selected anonymised lift profiles, monthly reach (Native 50-customer cohort, Q1 2026).

Several features of the profile distribution are worth noting. First, the lift ratios for the selected upper-tail cases range from roughly two times (Profile C, a high-baseline account that grew modestly in proportional terms but substantially in absolute terms) to roughly one hundred and eighty times (Profile J, a near-zero baseline that grew to mid-five-figure monthly reach). The wide range reflects the underlying mathematical sensitivity of ratio metrics to small denominators. Second, the high-lift cases are not predominantly the new-activation cases; substantial lift can occur from non-trivial pre-period baselines (Profiles A and B both had measurable pre-period reach and grew by an order of magnitude). Third, the absolute reach magnitudes post-Native are commercially meaningful: ten thousand to seventy thousand monthly reach for the cohort’s upper decile, in markets where the cohort businesses operate, represents a substantial proportion of the addressable audience.

We have chosen to present the upper-tail profiles rather than the median because the median experience is well-represented by the cohort averages reported in Table 6.1 (approximately five-fold reach lift, approximately ninefold posting lift). The variance information is most informative at the tails, where the distribution shape, in particular, the existence of a meaningful upper tail with order-of-magnitude lifts, has strategic implications for what AI-amplified marketing can deliver in the best case. We do not present the lower tail profiles in detail; the seven percent of accounts that did not improve on reach were discussed in section 6.2.

6.6 What the Data Cannot Tell Us

We close the findings chapter by being explicit about three things the cohort data cannot tell us, in the spirit of the methodological honesty introduced in chapter five.

First, the data cannot tell us what would have happened to the cohort businesses had they not adopted Native. The before-after design observes the actual trajectory, not the counterfactual. A causal estimate of the Native treatment effect would require either a randomised trial (which is operationally infeasible for a commercial subscription product) or a matched quasi-experimental design with similar non-adopting businesses (which is feasible and is the natural next study). We do not have that estimate.

Second, the data cannot tell us whether the lift is sustained beyond the post-period observation window. The longest post-period in the cohort is approximately four months. Whether the algorithmic platforms continue to reward the high-frequency, AI-amplified posting pattern over a one-year or three-year horizon is an empirical question that requires longer panels. We have early signals from the first wave of customers who have now been on the platform for several months, and those signals are broadly consistent with sustained lift, but the systematic longitudinal analysis is for a future paper.

Third, the data cannot tell us about downstream business outcomes. We measure posts, reach, and engagement, which are intermediate marketing outcomes. We do not measure sales, leads, conversion, customer lifetime value, or any of the financial outcomes that ultimately matter for the businesses in the cohort. The intermediate metrics are valuable proxies under most circumstances, but the proxy relationship can break down (for example, if the AI-amplified content reaches the wrong audience or produces the wrong type of engagement). We make no claim about downstream lift in this paper and we explicitly disclaim the inference.

With those limitations stated, we now turn to the discussion of what the findings imply for the broader theoretical and strategic questions the paper has set out to address.

7. Discussion

This chapter discusses the implications of the findings presented in chapter six and the framework developed in chapter four. We organise the discussion into five sections: reconciling the apparent contradiction between the performance and disclosure literatures; implications for the slop backlash discourse; implications for disclosure policy and regulation; implications for brand strategy and creative direction; and implications for the small and medium-sized businesses that are the primary commercial beneficiaries of AI-amplified marketing tools.

7.1 Reconciling the Performance and Disclosure Literatures

Sub-research question one asked how the performance findings from field experiments on AI advertising (Hartmann et al., 2025; Exner et al., 2025; Lee et al., 2025) can be reconciled with the trust-erosion findings from disclosure experiments (Schilke & Reimann, 2025; Koning & Voorveld, 2025). The findings of this paper, combined with the framework developed in chapter four, support the following reconciliation.

The performance literature and the disclosure literature are not in contradiction; they measure different things. The performance literature measures the behavioural response of audiences to ads they did not know were AI-generated. Hartmann et al. (2025), Exner et al. (2025), and Lee et al. (2025) all use real-world ad placements in which audience members were not informed that the creative work was AI-generated. The audiences clicked, viewed, and engaged based on the merits of the content as they perceived it. The merits were sufficient, on average, to produce equal or superior performance to human-only baselines. The disclosure literature, by contrast, measures the stated-preference response of audiences to ads they were explicitly told were AI-generated. When audiences are told, their persuasion knowledge is activated, their evaluative stance shifts, and they report less trust and less favourable attitudes.

Both findings are robust and both are correct within the conditions of their respective studies. The strategic question for brands and policymakers is which condition obtains in the wild. The empirical evidence suggests that the wild is overwhelmingly the no-disclosure condition: in the absence of regulatory or platform-mandated labelling, AI-generated content circulates in audience feeds without explicit attribution, and audiences engage with it as content rather than as AI content. This is the regime in which the performance literature is descriptive. The disclosure regime is the regime that would obtain if labelling requirements were comprehensively enforced; in that regime, the disclosure literature’s findings predict significant performance erosion.

Two further refinements deserve note. First, audiences’ ability to detect AI authorship without disclosure has improved over the period 2022–2026. Industry surveys report that approximately half of consumers now believe they can identify AI-generated content (Innovation Visual, 2025), and the visual heuristics audiences use have become well-known in popular discourse. The Exner et al. (2025) findings on the visual features that signal AI authorship, intense colour saturation, hyper-aestheticised composition, suspicious perfection, describe a detection capability that audiences are actively developing. The strategic consequence is that the no-disclosure regime is gradually being eroded not by regulation but by audience media literacy. AI-generated content that looks AI-generated is increasingly subject to the same trust penalty as disclosed AI-generated content, even without explicit disclosure.

Second, the trust penalty applies most strongly to the content categories in which audiences expect human authenticity to be material. Marketing imagery for products is one of these categories, but the strength of the expectation varies by sub-category. Hartmann et al. (2025) note that AI generation is less penalised in product visualisation contexts (where the audience expects post-production polish anyway) than in spokesperson contexts (where the audience expects human authenticity). The Taste Layer framework predicts this variation: the more the content invites parasocial relationship with the brand, the more the audience values evidence of human authorship, and the more carefully the platform must design output for the category. A platform delivering taste in product imagery is solving a different problem than a platform delivering taste in spokesperson video, and good platform design recognises the distinction.

Kirk and Givi (2025) sharpen this picture experimentally. Across seven preregistered studies of AI-generated marketing communications, the authorship penalty concentrates in emotional rather than factual messages, is serially mediated by perceived authenticity and moral disgust, attenuates when AI only edits rather than authors the communication, and reverses entirely when the comparison baseline is reused rather than original human content. The moderation pattern matters as much as the main effect: it identifies emotional register, not AI involvement as such, as the variable that gates the penalty, and it establishes that the baseline against which AI content is judged is the realistic alternative, not an idealised artisanal human. For the SMB segment studied in this paper, whose realistic counterfactual is sporadic or recycled content rather than commissioned craft, the relevant baseline comparison is the one under which the penalty disappears.

The cohort data presented in chapter six is consistent with this reconciliation. The cohort businesses are operating in the no-disclosure regime, their AI-amplified content is reaching and engaging audiences at multiples of pre-Native baselines, and the engagement-per-impression rate is preserved, indicating that the content is being received as content rather than penalised as AI content. The cohort data does not directly speak to what would happen if these accounts began making prominent AI-disclosure statements in every post. Based on the disclosure literature, we would predict significant performance erosion from prominent disclosure, and we would recommend against it on strategic grounds. We return to the policy implications of this position below.

7.2 Implications for the Slop Backlash

Sub-research question four asked how the slop discourse documented in trade press and cultural commentary maps onto the distinct phenomena of engagement farming, brand-led cost-cutting, and population-level content homogenisation. The Taste Layer framework provides a clean answer.

Engagement farming is a failure of the entire pipeline. There is no signal extraction in the meaningful sense (no real business to extract from), no curation gate, and no genuine audience-value intent. The AI execution is configured to trigger algorithmic reward functions rather than to serve audience interest. The output is unambiguously slop on any definition of the term. The remedy is not better AI but platform-side content policy that suppresses or demonetises engagement-farming accounts. This is a policy problem for the platforms, not an opportunity for tools like Native to compete; engagement farms do not buy marketing tools, they exploit free generative APIs. The slop discourse is correct that engagement farming is degrading public discourse, but it is incorrect to attribute the phenomenon to legitimate AI-amplified marketing.

Brand-led cost-cutting is a failure of the curation gate. The Coca-Cola, McDonald’s, and J.Crew cases discussed in section 2.2 all involve brands using AI to reduce the cost of campaigns that previously required commissioned creative work, without preserving the editorial layer that would have caught the obvious problems. The failure is identifiable by characteristic signatures: hands with extra fingers, animals out of season, lighting inconsistencies, generic compositions, brand voice violations. These are not edge cases of generative model capability; they are routinely catchable by a competent reviewer. The fact that they were not caught indicates that the curation gate had been removed from the workflow, presumably for cost reasons. The remedy is for brands to restore the curation gate, recognising that the value of AI amplification is the cost reduction of production while preserving the curation function, not the elimination of both.

Population-level content homogenisation is the deepest of the three problems and the one for which there is no clean operational remedy at the individual brand level. If many brands use overlapping generative models with overlapping training data, output convergence is the predictable result. Doshi et al. (2024) document this phenomenon experimentally in creative writing. The cohort data presented in this paper cannot directly speak to homogenisation at the population level because we observe individual account trajectories, not the aggregate diversity of the content ecosystem. We acknowledge the concern honestly. The defence at the individual brand level is investment in the signal layer (stage one of the Taste Layer framework): brands whose underlying business has distinctive character, a particular market, a particular way of speaking, a particular set of products that nobody else sells, produce more distinctive output than generic businesses, even from the same underlying models. For the platform operating on behalf of the brand, the operational implication is that signal extraction must be specific enough to capture what is distinctive about each customer rather than collapsing every customer toward a generic SMB voice. The defence at the platform level beyond this is more challenging and may require platform-side investment in diversity-preserving distribution algorithms, a research direction that is beginning to receive attention (Liu et al., 2025).

The three failure modes have very different policy and operational implications. Conflating them under the slop label, as much of the trade press discourse does, produces strategic and regulatory confusion. The most important consequence of this conflation is that legitimate AI-amplified marketing, the kind documented in the cohort data presented in this paper, gets rhetorically positioned alongside engagement farming and brand cost-cutting failures, and is then subjected to the same regulatory and reputational pressures. We argue this conflation is analytically incorrect and strategically counterproductive.

A detail from the 2026 listening data illustrates how much framing, rather than technology, drives the backlash. In Brand24's May 2026 corpus, the hashtag aigenerated was used in an almost entirely positive register while the hashtag aislop was almost entirely negative, applied to the same underlying category of content (Brand24, 2026). Audiences are not sorting content by production method; they are sorting it by perceived care, exactly as the effort-heuristic account in chapter three predicts. The same period produced the first wave of explicitly anti-AI brand positioning in mainstream advertising, with several major consumer brands running campaigns in early 2026 built on the promise of human-made creative. The framework of this paper predicts that such positioning works precisely as long as it is operationally true, and becomes a liability the moment audiences detect the gap between the claim and the production pipeline.

7.3 Implications for Disclosure Policy and Regulation

The disclosure paradox documented in Schilke and Reimann (2025) and Koning and Voorveld (2025) creates a genuine tension for regulators. The European Union’s Digital Services Act and the emerging AI Act provisions on transparency assume that disclosure protects consumers. The empirical record suggests that prominent disclosure damages the disclosing party in ways that may reduce the supply of legitimate AI-amplified content available to consumers, while doing little to address the actual harms of engagement farming and brand cost-cutting failures.

We do not take a strong policy position in this paper on whether disclosure should be required. We do offer two observations that may be useful for the policy discussion.

First, the consumer protection that disclosure is intended to provide can be substantially achieved through means that do not activate the persuasion knowledge penalty. Bekken’s (2025) participants converged on a specific suggestion that has not received adequate attention in the policy literature: disclosure in alt text, in fine print on brand websites, or in metadata accessible on demand rather than prominently in the content itself. This approach satisfies the audience right to information about content provenance for those audience members who actively seek that information, while not activating persuasion knowledge for the broader audience that does not. From a consumer protection standpoint, this is arguably superior to prominent labelling: audiences who care about AI provenance can find the information, and audiences who do not care are not subjected to a paternalistic warning that distorts their engagement.

Second, the most direct consumer protection in AI-amplified marketing is not against AI use per se but against the failure modes that AI use enables. Brand cost-cutting failures harm consumers by exposing them to low-quality, sometimes deceptive content. Engagement farming harms consumers by degrading the information value of platform feeds. Population-level homogenisation harms consumers by reducing the diversity of cultural production. Each of these harms admits regulatory remedies that are more targeted than blanket disclosure requirements: platform content quality policies for engagement farming, advertising standards enforcement for cost-cutting failures, and competition or industrial policy interventions for homogenisation. A disclosure-only regulatory regime addresses the visible signal of AI use without addressing the underlying harms.

We do not argue that disclosure should never be required. There are specific contexts, political advertising, medical claims, image manipulation in news contexts, in which prominent disclosure is justified by considerations that override the engagement penalty. We argue only that the assumption of disclosure as a default protective measure deserves more careful empirical scrutiny than it has so far received, and that the Schilke and Reimann findings have substantial policy implications that the current EU framework does not adequately incorporate.

Recent experimental work adds a nuance that policymakers should weigh. Shi and Jiang (2026) demonstrate across three experiments that AI disclosure labels operate as a double-edged sword: labels increase perceived novelty, which improves advertisement and product attitudes, while simultaneously reducing perceived authenticity, which damages them, and the net effect depends on the consumer's regulatory focus. The finding does not overturn the disclosure penalty documented above, but it does suggest the penalty is a net effect of competing paths rather than a uniform aversion, and that its size will vary across audiences and framings. The policy implication is unchanged in direction, comply where required and do not gold-plate, but the research frontier is clearly moving toward identifying the conditions under which disclosure is least costly.

7.4 Implications for Brand Strategy and Creative Direction

For brands operating in the AI-amplified marketing environment, the Taste Layer framework yields several strategic implications worth making explicit.

Invest in the signal, not the model. The competitive advantage in AI-amplified marketing is at the input boundary. Strong signal produces strong output regardless of which generative model is used; weak signal produces weak output regardless. The implication varies by firm size. For large brands with internal marketing teams, the functions worth staffing internally are creative direction, brand voice development, and audience insight, the functions that produce a strong signal in the form of an explicit brief. For small and medium-sized businesses without internal marketing teams, the equivalent is choosing tools that extract the signal effectively from minimal customer input, and being a sufficiently coherent business that the signal is recoverable. The common mistake brands have made over the 2024–2026 period is investing in AI tooling without correspondingly investing in the signal-quality side of the equation. Hartmann et al. (2025) demonstrate that the best AI execution exceeds professional human creative work; the signal is what makes that possible, whether the signal is supplied actively by a creative director or extracted passively from a business’s existing presence.

Invest in the curation gate. The most consequential operational decision for a brand using AI-amplified marketing is who decides what gets shipped. The Coca-Cola, McDonald’s, and J.Crew cases discussed in section 2.2 are all curation gate failures. The brand had access to the same generative tools that other brands had; the difference is that the editorial layer was removed or attenuated, and the obvious problems with the output were not caught. The curation gate need not be expensive in absolute terms, a single editor reviewing the day’s output at a small business is sufficient, but it must be present. Brands that treat the curation gate as a cost to be eliminated are brands that will produce slop, regardless of how sophisticated their AI tooling is. The recommendation is mechanism-backed rather than folk wisdom: the gate exists to manage the effort-cue density on which audience valuation runs (Kruger et al., 2004) and to keep emotional-register content out of the uncanny zone where experience claims meet perceived machine authorship (Gray & Wegner, 2012; Kirk & Givi, 2025). Section 4.7.4 demonstrates that the gate can be instrumented and measured rather than merely staffed: acceptance rate is the gate expressed as a number.

Aesthetic strategy as anti-detection strategy. The Exner et al. (2025) findings on the visual features that signal AI authorship provide a near-operational checklist for brands. Hyper-saturated colour palettes, glossy hyper-realism, suspiciously perfect composition, and stock-photo-style generic framing are all penalised. Brands should actively design against these features. The most successful AI-amplified visual marketing during 2024–2026 has tended to look hand-made, slightly imperfect, textured, and culturally specific, not because AI cannot produce other styles but because audiences associate these styles with human creative work and reward them accordingly. The strategic implication is that brand visual identity is now also an AI-detection-avoidance strategy, and brands without a strong visual identity (defaulting to the generative model’s aesthetic defaults) are at a structural disadvantage.

Frequency without homogenisation. The cohort data shows that AI amplification produces large frequency lifts. Frequency is operationally valuable but introduces the risk of homogenisation: a brand publishing nine times more content per month than it did before may publish nine times more on-brand content or it may publish nine times more generic content. The latter risk is real and growing as more brands adopt similar AI tools. The defence is in the strength of the input signal (the distinctiveness of the brand as it actually is, and how well that distinctiveness gets extracted into the generation process) and in the cadence-shaping decisions about what mix of content types is published. Brands that publish high frequency with high consistency in voice and perspective will increasingly outcompete brands that publish high frequency with low consistency, because the audience is increasingly capable of distinguishing the two.

Treat the brand voice as the disclosure. Schilke and Reimann (2025) demonstrate that explicit AI disclosure erodes trust regardless of underlying quality. The practical alternative is to let taste do the work that disclosure cannot: a brand voice that is consistent, distinctive, culturally specific, and editorially disciplined communicates competence and care without needing to make claims about who or what produced it. Audiences engaging with content that has these characteristics extend trust based on what they see, not based on what the platform tells them about itself. The brand voice is the disclosure. The taste is the disclosure. The strategic alignment is between the platform’s actual operational reality (taste is encoded in the product; the platform performs the work; the output reflects the brand) and the brand’s public signal.

7.5 Implications for Small and Medium-Sized Businesses

The cohort presented in this paper consists of small and medium-sized businesses, and the strategic implications most relevant to that segment deserve particular attention. SMBs have historically operated under a structural disadvantage in marketing: the production cost of professional marketing content has been roughly fixed across firm sizes (the cost of producing a quality social post is roughly the same whether the firm has ten employees or ten thousand), while the revenue available to amortise that cost has been very different. The result has been a marketing-capability gap, with large firms publishing at scale and small firms publishing sporadically or not at all.

AI amplification substantially narrows this gap. The cohort data presented in chapter six shows that small businesses adopting AI-amplified marketing can move from sporadic publication (or no publication) to consistent publication, with corresponding lifts in reach and engagement. This is operationally significant. It is also, we believe, socially significant: a marketing environment in which only large firms can afford to participate is an environment in which consumer information about small-firm alternatives is suppressed. AI amplification, used through tools that preserve the curation gate, is a partial corrective to this asymmetry.

Three operational observations follow for SMBs evaluating AI-amplified marketing tools. First, the most important property of a tool is not how sophisticated its generation capability is, capabilities have largely converged across competing platforms, but how well the tool extracts signal from minimal customer input and how aggressively the tool filters its own output. Tools that publish output without any filtration produce slop regardless of how good the underlying generation is. Tools that filter aggressively, whether through internal editorial layers, automated quality checks, or hybrid systems, produce output that competes credibly with professional creative work. The customer’s evaluation criterion should be “does the output look like something my business would actually publish” rather than “does the tool give me lots of options to direct the work.” Second, the customer’s active contribution can be minimal but cannot be zero. Some level of feedback, occasional approval, occasional rejection, occasional correction of brand-voice drift, is what keeps the signal current as the business evolves. A customer who completely disengages from the loop will, over time, drift away from their own brand as the platform’s signal-extraction model becomes stale. Third, do not confuse AI amplification with the elimination of the marketing function. The marketing function is still being performed; the question is who performs it. Tools that perform it well on the customer’s behalf are valuable; tools that pretend the function does not exist produce slop.

We note that the cohort data shows substantial variance across accounts (section 6.5), and we are not arguing that AI amplification produces uniform success. The accounts that benefited most were those with the most room to grow (low pre-Native baselines) and those whose underlying business had recoverable distinctive character that the platform could extract and amplify (specific products, specific markets, specific ways of speaking). The accounts that benefited least were those with already-high baselines, those whose underlying business was indistinguishable from many others in their category (low signal recoverability), or those operating in markets where the social media platforms themselves are not the dominant channel for audience reach. The framework is not magic; the lift available depends on what there is to amplify in the first place.

8. Conclusion

This paper has examined the apparent contradiction between two robust empirical findings about generative AI in marketing: that AI-amplified marketing content can match and frequently exceed human-only content on field-performance metrics, and that audiences punish content perceived or disclosed as AI-generated with measurable trust erosion. We have argued, drawing on four research streams in marketing and communication theory and on observational data from a cohort of fifty small and medium-sized businesses using an AI-amplification platform, that the contradiction is resolvable through a framework we call the Taste Layer. This concluding chapter summarises the contributions of the paper, states its limitations candidly, and offers directions for future research.

8.1 Summary of Contributions

The paper makes four contributions to the academic literature and to marketing practice.

Theoretical synthesis. We synthesise four research streams, audience value perception, persuasion knowledge and disclosure trust, the Hartmann/Exner field experimental work, and the human–AI collaboration literature, into a unified framework that articulates the conditions under which AI-amplified marketing produces audience approval rather than backlash. The Taste Layer framework locates competitive advantage in the input boundary (signal extraction) and the output boundary (curation gate) rather than in the AI execution layer, which has been commoditised across 2024–2026. The framework provides a parsimonious explanation for the apparent contradiction in the literature: audiences punish content perceived as AI-generated, but the most consequential perception is not driven by the underlying production method but by the visual and stylistic signals that the production process leaves behind. Brands whose pipelines preserve taste judgment at both boundaries avoid these signals; brands whose pipelines do not, do not. A practical refinement of the framework, important for the SMB context, is that the taste functions need not be performed by the customer; they can be performed by the platform on the customer’s behalf, provided the platform is designed to do so.

Empirical contribution. We present the first systematic operational data from a real-world cohort of small and medium-sized businesses using an AI-native marketing platform. The cohort data shows substantial lifts in posting frequency (approximately ninefold), monthly organic reach (approximately fivefold), and monthly engagement (approximately fivefold), with ninety-three percent of accounts increasing reach and ninety-six percent increasing posting frequency. Nine of fifty accounts transitioned from inactive to consistently active publication. The lift magnitudes are consistent with the theoretical prediction that platform-encoded taste, operating in a delegation regime where the customer is in the outcome loop rather than the production loop, can deliver the field-experimental performance documented in Hartmann et al. (2025), Exner et al. (2025), and Lee et al. (2025). We are explicit throughout about the observational rather than causal nature of these findings.

Disaggregation of the slop discourse. We argue that the AI slop discourse conflates three analytically distinct phenomena (engagement farming, brand-led cost-cutting, and population-level homogenisation) and that this conflation produces strategic and regulatory confusion. The three phenomena have different causes, different remedies, and different relationships to legitimate AI-amplified marketing. Engagement farming is a platform content policy problem. Brand cost-cutting is a curation gate failure addressable by editorial discipline. Homogenisation is a structural property of large-scale AI adoption that requires investment in signal quality at the firm level (and signal-extraction quality at the platform level) and may require platform-side diversity-preserving distribution interventions at the system level. Treating these as a single phenomenon, as much trade press discourse does, obscures all three.

Policy contribution. We offer a contribution to the disclosure policy debate by integrating the Schilke and Reimann (2025) findings with the practical insight, surfaced inductively in Bekken (2025), that audiences may accept lower-prominence disclosure mechanisms (alt text, fine print, on-demand metadata) without the trust penalty associated with prominent disclosure. This positions disclosure policy as a more textured design problem than the current EU framework treats it, and suggests that consumer protection in AI-mediated marketing may be better served by targeted interventions on the underlying failure modes than by blanket prominent labelling.

8.2 Limitations and Future Research

The limitations of the paper are substantial and worth restating explicitly. The cohort size is modest (n = 50), the cohort is geographically concentrated in Norway and Northern Europe (eighty-six percent), the design is observational rather than experimental, and the analysis was conducted internally rather than by independent auditors. Customers are self-selected into the platform, the cohort is biased toward retained customers (survivor bias), the post-Native trajectory is confounded with other marketing investments the businesses made over the same period, and the longest observation window is approximately four months. These limitations mean the cohort findings should be read as evidence about what self-selected adopters experience, not as causal estimates of treatment effects on a counterfactual non-adopter.

Six directions for future research follow directly from these limitations.

Matched quasi-experimental analysis. The most direct strengthening of the empirical contribution would be a matched-pair design in which Native customers are matched to similar non-customers on pre-period industry, geography, size, and performance. This would substantially improve causal identification without requiring a randomised trial. The matching data is in principle available from public platform analytics, and we anticipate publishing this analysis in a follow-up paper.

Longitudinal extension. The current observation window of approximately four months is short. The most important open empirical question is whether the lift is sustained over longer horizons (one year, three years) or whether it attenuates as the platforms’ algorithmic dynamics adjust to high-frequency AI-amplified content. We have preliminary signals from the first wave of customers with longer tenure, and these will be the subject of a future paper as the data matures.

Downstream business outcomes. We measure intermediate marketing outcomes (posts, reach, engagement). The translation from these intermediate metrics to financial outcomes (sales, leads, customer acquisition cost, lifetime value) is the question that ultimately matters for the businesses adopting the platform. Linking platform performance data to customer-side business outcomes is a research design that requires substantial customer cooperation and is not feasible at scale within the current data infrastructure. Targeted case studies with willing customers, however, are feasible and would strengthen the operational claims.

Diversity collapse measurement. The most consequential open question about AI-amplified marketing at the population level is whether and how much it produces convergence in style, voice, and content across brands. Doshi et al. (2024) document the phenomenon in creative writing experimentally; the analogous analysis in visual marketing content is much more challenging because the content is multimodal and the population is larger. Methodological work on operationalising content diversity for marketing visual ecosystems is needed.

Cross-cultural extension. The cohort presented in this paper is concentrated in Norway and Northern Europe. The Bekken (2025) thesis raised concerns about the generalisability of findings from Nordic samples, given the unusually high media literacy and platform-trust characteristics of those populations. Replication of both the qualitative and the operational findings in other geographic settings (North America, Southeast Asia, Latin America) would speak directly to the external validity of the Taste Layer framework. Whether the framework operates similarly in markets with different platform dynamics, different audience media literacy norms, and different cultural relationships to AI is open.

Disclosure design experiments. The policy contribution of this paper rests on the empirical claim that lower-prominence disclosure mechanisms can satisfy audience right-to-know interests without activating the persuasion knowledge penalty. This is a plausible inference from the existing literature but it has not been tested experimentally in the marketing context. A pre-registered between-subjects experiment comparing prominent disclosure, lower-prominence disclosure (alt text, fine print), and no disclosure across matched content would directly inform the disclosure policy discussion. We anticipate that such an experiment will be conducted by the academic community and would welcome the data it produces.

Finally, the research base itself is maturing in ways that will discipline future versions of this argument. The first systematic review of consumer trust in AI-generated marketing content (Baryshkov, 2026) now exists, and web-scale measurement of AI content share has moved from extrapolation to longitudinal observation (Axios, 2026). Both developments invite a more precise research agenda: field replication of the authenticity-mediation structure in organic social content rather than paid advertising, where nearly all existing experiments are situated, and longitudinal tracking of whether the plateau in AI-generated article share holds as agentic production tools mature.

8.3 Closing Remarks

The intersection of generative AI and marketing communications is at an unusual moment. The academic empirical literature, the trade press discourse, and the operational practice of running marketing tools are giving substantively different answers to the same question. The peer-reviewed field experiments document performance gains from AI-amplified marketing that exceed any comparable technology introduction in the digital marketing era. The trade press documents cultural backlash against AI-generated content with increasing intensity, including the elevation of slop to Word of the Year status. The operational practice, in our experience and in the cohort data we present, is consistent with neither pure narrative: it shows substantial gains for businesses that adopt AI amplification thoughtfully, while also containing the visible failure cases that fuel the backlash.

Our central argument is that the apparent contradictions in this picture dissolve when one distinguishes the AI execution layer from the taste judgment layers that bracket it. The execution layer has been commoditised and its capabilities now exceed human creative work on many metrics that matter to advertisers. The judgment layers, signal extraction at the front and the curation gate at the back, are not commodities. They are functions of taste, brand-specific perspective, and editorial discipline, and they can be performed by the customer, by the platform on the customer’s behalf, or by some combination of the two. The brands and businesses whose pipelines preserve these functions benefit from AI amplification at the magnitudes documented in the field-experimental literature and reflected in our cohort data. The brands and businesses whose pipelines do not produce the failures documented in the slop discourse. The pattern is consistent across the evidence.

If this analysis is correct, two broad implications follow. For practitioners, the operational mandate is to invest in taste as a pipeline property rather than as an organisational role: large brands invest in internal taste capability through creative direction; small brands choose tools that perform taste functions on their behalf; in both cases, the work gets done, but the question of who does it is determined by firm size and operating model rather than by the framework. For policymakers and platforms, the regulatory mandate is to address the specific failure modes that AI use enables rather than AI use itself, because broad-brush disclosure or restriction regimes are likely to suppress legitimate AI-amplified content without addressing the actual harms. For researchers, the methodological mandate is to continue separating the perceptual, behavioural, and operational dimensions of the question, because the answers differ across dimensions and conflating them obscures the underlying structure.

We close with a note on the institutional position from which this paper is written. Native is a commercial firm that benefits financially from the adoption of AI-amplified marketing tools. We have endeavoured to write this paper as research rather than as marketing material, and we have presented data and theoretical arguments that have implications we are not entirely comfortable with as a commercial operator (most notably the diversity collapse concern, which is a long-run risk to the category we operate in). We invite criticism and disagreement. The framework presented here is offered as a contribution to a conversation, not as a settled position. The empirical landscape will continue to evolve, and we expect the conclusions reached here to be revised in light of evidence we have not yet seen. What we offer in May 2026 is a synthesis of what we currently understand, presented with the limitations we are aware of stated explicitly. We hope it is useful.

References

References are formatted in APA 7th-edition style. Where peer-reviewed publication details are available, they are cited; where work is at the working-paper or preprint stage, the SSRN or institutional identifier is provided. Industry sources are cited with the publishing organisation and date.

Ahrefs. (2025, May 19). 74% of new webpages include AI content: A study of 900,000 pages. Ahrefs Blog. https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/

Alter, A. L., & Oppenheimer, D. M. (2009). Uniting the tribes of fluency to form a metacognitive nation. Personality and Social Psychology Review, 13(3), 219–235.

Axios. (2026, May 15). How much of the web is written by AI? [Analysis of Graphite data]. Axios. https://www.axios.com/2026/05/15/human-vs-ai-written-articles

Bain & Company. (2025, July 1). Agentic AI commerce hinges on consumer trust. Bain Insights. https://www.bain.com/insights/agentic-ai-commerce-hinges-on-consumer-trust/

Baryshkov, K. (2026). Consumer trust in AI-generated marketing content: A systematic literature review and research agenda. American Impact Review. https://doi.org/10.66308/air.e2026024

Bekken, B. A. (2025). What is engaging content? A study of AI value perception within digital marketing [Master’s thesis, Norwegian University of Science and Technology]. NTNU Open Archive.

Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of Marketing Research, 49(2), 192–205.

Borkakoty, K. (2025, February 2). Social media metrics: All types explained [2025]. Sprinklr Blog. https://www.sprinklr.com/blog/social-media-metrics/

Brand24. (2026, June 12). Why do people hate AI content? [2026 research]. Brand24 Blog. https://brand24.com/blog/why-people-hate-ai-content-report/

Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(5), 809–825.

Chen, H., Chan-Olmsted, S., Kim, J., & Sanabria, I. M. (2021). Consumers’ perception on artificial intelligence applications in marketing communication. Qualitative Market Research, 25(1), 125–142.

Click Consult. (2024, December). New research sheds insight on customer perception on the use of AI in marketing activity. Click Consult Insights. https://www.click.co.uk/blog/research-customer-perception-ai-marketing/

CNN. (2025, December 16). Merriam-Webster’s 2025 word of the year takes aim at poor AI content. CNN Business. https://www.cnn.com/2025/12/16/tech/slop-merriam-webster-2025

Contentgrip. (2025, November 12). Coca-Cola’s AI holiday ad triggers backlash. Content Grip. https://www.contentgrip.com/coca-cola-ai-holiday-ad-backlash/

Cool Down. (2026, March 20). Co-creator behind divisive Coca-Cola ad doubles down: ‘I’m super proud of that ad.’ The Cool Down. https://www.thecooldown.com/

Cro Metrics. (2026, January). AI and marketing predictions 2026: Customer journeys split as consumers decide between AI delegation and authentic personal experience [Industry report]. https://www.businesswire.com/

Davenport, T., Guha, A., Grewal, D., & Bressgott, T. (2020). How artificial intelligence will change the future of marketing. Journal of the Academy of Marketing Science, 48(1), 24–42.

Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2025). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper No. 24-013.

Deloitte. (2026, February 24). Agentic commerce: AI shopping agents guide 2026 [Industry research report]. Deloitte Insights. https://www.deloitte.com/us/en/Industries/consumer/articles/agentic-commerce-ai-shopping-agents-guide.html

Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126.

Doshi, A. R., Hauser, O. P., & Mollick, E. R. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290

EMarketer. (2025, October). Marketers insist on ethics as AI agents reshape shopping. EMarketer Insider Intelligence. https://www.emarketer.com/content/marketers-insist-on-ethics-ai-agents-reshape-shopping

EMarketer. (2026, January). As ad platforms launch agentic AI features, marketer distrust persists. EMarketer Insider Intelligence. https://www.emarketer.com/content/ad-platforms-launch-agentic-ai-features–marketer-distrust-persists

Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886.

Exner, Y., Hartmann, J., Netzer, O., Zhang, S., & Ding, Z. (2025). AI in disguise: How AI-generated ads’ visual cues shape consumer perception and performance [Working paper]. Columbia Business School Research Paper. https://ssrn.com/abstract=5096969

EY. (2025). Consumer AI trust survey: Global findings on autonomous AI adoption [Industry research report]. Ernst & Young.

Eyal, N. (2014). Hooked: How to build habit-forming products. Portfolio/Penguin.

Ferster, C. B., & Skinner, B. F. (1957). Schedules of reinforcement. Appleton-Century-Crofts.

Festinger, L. (1954). A theory of social comparison processes. Human Relations, 7(2), 117–140.

Friestad, M., & Wright, P. (1994). The persuasion knowledge model: How people cope with persuasion attempts. Journal of Consumer Research, 21(1), 1–31.

Gray, K., & Wegner, D. M. (2012). Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition, 125(1), 125–130.

Grewal, D., Satornino, C. B., Davenport, T., & Guha, A. (2025). How generative AI is shaping the future of marketing. Journal of the Academy of Marketing Science, 53, 702–722. https://doi.org/10.1007/s11747-024-01064-3

Grigsby, J. L., Michelsen, M., & Zamudio, C. (2025). Service ads in the era of generative AI: Disclosures, trust, and intangibility. Journal of Retailing and Consumer Services, 87, 104427.

Hartmann, J., Exner, Y., & Domdey, S. (2025). The power of generative marketing: Can generative AI create superhuman visual marketing content? International Journal of Research in Marketing, 42(1), 13–31. https://doi.org/10.1016/j.ijresmar.2024.09.002

Hermann, E., & Puntoni, S. (2025). Generative AI in marketing and principles for ethical design and deployment. Journal of Public Policy & Marketing, 44(3), 332–349. https://doi.org/10.1177/07439156241309874

Hoechsmann, M., & Poyntz, S. R. (2012). Media literacies: A critical introduction. Wiley-Blackwell.

Holmström, J., & Carroll, N. (2024). How organizations can innovate with generative AI. Business Horizons.

Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction: Observations on intimacy at a distance. Psychiatry, 19(3), 215–229.

IBTimes UK. (2025, November 24). Coca-Cola faces intense backlash for using AI again for its 2025 Christmas ad: ‘Disgusting.’ International Business Times. https://www.ibtimes.co.uk/

Innovation Visual. (2025). AI content detection and consumer awareness: Annual industry survey. Innovation Visual Insights.

Jago, A. S. (2019). Algorithms and authenticity. Academy of Management Discoveries, 5(1), 38–56.

Katz, E., Blumler, J. G., & Gurevitch, M. (1973). Uses and gratifications research. Public Opinion Quarterly, 37(4), 509–523.

Kirk, C. P., & Givi, J. (2025). The AI-authorship effect: Understanding authenticity, moral disgust, and consumer responses to AI-generated marketing communications. Journal of Business Research, 186, 114984.

Koning, M., & Voorveld, H. A. M. (2025). Disclosing AI in advertising: Effects on persuasion knowledge, trust, and advertising effectiveness. Journal of Marketing Communications. Advance online publication.

Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. (2014). Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences, 111(24), 8788–8790.

Kruger, J., Wirtz, D., Van Boven, L., & Altermatt, T. W. (2004). The effort heuristic. Journal of Experimental Social Psychology, 40(1), 91–98.

Kumar, V., Ashraf, A. R., & Nadeem, W. (2024). AI-powered marketing: What, where, and how? International Journal of Information Management, 77, 102783.

Lee, S., Todri, V., Adamopoulos, P., & Ghose, A. (2025). The impact of visual generative AI on advertising effectiveness [Working paper]. Emory University Goizueta Business School. https://ssrn.com/abstract=4895789

Liu, X., Sun, T., Goli, A., & Netzer, O. (2025). Human–AI collaboration in social media advertising: A large-scale field experiment. Marketing Science. Advance online publication.

Livingstone, S. (2004). Media literacy and the challenge of new information and communication technologies. The Communication Review, 7(1), 3–14.

Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103.

Macready, H. (2024, April 15). Organic vs. paid social media: Differences and strategic tips. Hootsuite Blog. https://blog.hootsuite.com/organic-vs-paid-social-media/

Mariani, M. M., Perez-Vega, R., & Wirtz, J. (2022). AI in marketing, consumer research and psychology: A systematic literature review and research agenda. Psychology & Marketing, 39(4), 755–776.

McKinsey & Company. (2026, January 28). The automation curve in agentic commerce [Industry report]. McKinsey QuantumBlack. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-automation-curve-in-agentic-commerce

McKinsey & Company. (2026, March 2). Europe’s agentic commerce moment: Decision influence is here, execution is coming [Industry report]. McKinsey QuantumBlack.

Meltwater & Brandwatch. (2025). Annual sentiment review: AI slop discourse 2024–2025 [Industry report].

Merriam-Webster. (2025, December 15). Word of the Year 2025: Slop. Merriam-Webster Wordplay. https://www.merriam-webster.com/wordplay/word-of-the-year

Metzger, M. J., & Flanagin, A. J. (2013). Credibility and trust of information in online environments: The use of cognitive heuristics. Journal of Pragmatics, 59, 210–220.

NBC News. (2024, November 18). Coca-Cola causes controversy with AI-generated ad. NBC News. https://www.nbcnews.com/tech/innovation/coca-cola-causes-controversy-ai-made-ad-rcna180665

Newman, G. E., & Bloom, P. (2012). Art and authenticity: The importance of originals in judgments of value. Journal of Experimental Psychology: General, 141(3), 558–569.

Nightingale, S. J., & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences, 119(8), e2120481119.

Nuremberg Institute for Market Decisions (NIM). (2025). The AI transparency dilemma: When disclosure damages brands [Industry research report]. NIM.

OpenAI. (2022, November 30). Introducing ChatGPT. OpenAI Blog. https://openai.com/index/chatgpt/

Pew Research Center. (2025). Public attitudes toward artificial intelligence across 25 countries [Survey report]. Pew Research Center.

Puntoni, S. (2024). Artificial intelligence in marketing: From computer science to social science. Journal of Macromarketing, 44(4), 883–885.

Puntoni, S., Reczek, R. W., Giesler, M., & Botti, S. (2021). Consumers and artificial intelligence: An experiential perspective. Journal of Marketing, 85(1), 131–151.

Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking.

Saha, T. (2024). AI and content in marketing: Correlations on perception. International Journal for Multidisciplinary Research, 6(3), 1–40.

Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. https://doi.org/10.1016/j.obhdp.2025.104405

Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and quasi-experimental designs for generalized causal inference. Houghton Mifflin.

Search Engine Land. (2026, June). AI search adoption rises as consumer trust declines: Study. Search Engine Land. https://searchengineland.com/ai-search-adoption-rises-consumer-trust-declines-study-480338

Shi, Y., & Jiang, Z. (2026). Consumer responses to AI disclosure labels: The role of novelty and authenticity. SAGE Open. https://doi.org/10.1177/21582440261417793

Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger (Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press.

Stanford HAI. (2025). Artificial Intelligence Index Report 2025. Stanford Institute for Human-Centered AI.

Sundar, S. S., & Limperos, A. M. (2013). Uses and grats 2.0: New gratifications for new media. Journal of Broadcasting & Electronic Media, 57(4), 504–525.

Teepapal, T. (2025). AI-driven personalization: Unraveling consumer perceptions in social media marketing. Computers in Human Behavior, 162, 108234.

Toff, B., & Simon, F. M. (2024). ‘Or they could just not use it?’: The dilemma of AI disclosure for audience trust in news. The International Journal of Press/Politics. Advance online publication.

Wilson, H. J., & Daugherty, P. R. (2018). Collaborative intelligence: Humans and AI are joining forces. Harvard Business Review, 96(4), 114–123.

Yahoo Finance. (2026, April 19). AI traffic to US retailers jumps 393% in Q1 as agentic shoppers outspend humans. Yahoo Finance. https://finance.yahoo.com/sectors/technology/articles/ai-traffic-us-retailers-jumps-160141756.html

Zajonc, R. B. (1968). Attitudinal effects of mere exposure. Journal of Personality and Social Psychology, 9(2, Pt. 2), 1–27.

Zeithaml, V. A. (1988). Consumer perceptions of price, quality, and value: A means-end model and synthesis of evidence. Journal of Marketing, 52(3), 2–22.

Appendix A: Cohort Composition Summary

This appendix provides additional descriptive detail on the fifty-customer cohort analysed in chapters five and six, beyond what is presented in Table 5.1. The information here is intended to support the interpretation of the headline findings without introducing identifiable customer detail.

A.1 Tenure distribution

Tenure on the Native platform at the cohort-selection date (1 April 2026) ranged from 60 days (the minimum tenure required by the cohort criteria) to approximately 380 days, with a median tenure of approximately 145 days. Approximately sixty percent of accounts had been on the platform for between 90 and 180 days, with the remainder distributed across shorter (60–90 day) and longer (180–380 day) tenures. The composition reflects the platform’s growth trajectory over the period: most cohort accounts joined Native during the November 2025 through January 2026 expansion window.

A.2 Subscription tier distribution

Native offers subscription tiers that vary in publishing cadence, number of platforms connected, and feature access. The cohort composition by tier at the cohort-selection date was approximately balanced, with no single tier representing more than forty percent of the cohort. The variance in publishing cadence across tiers is reflected in the per-account posting frequency lift figures reported in chapter six; accounts on higher-cadence tiers contribute disproportionately to the headline nine-fold lift average, while accounts on lower-cadence tiers contribute disproportionately to the reach efficiency observed in the ratio of reach lift to post lift.

A.3 Platform mix

The cohort published across an average of approximately four social media platforms per account. The most common platforms were Meta (Instagram and Facebook combined, approximately ninety percent of accounts), LinkedIn (approximately seventy percent), TikTok (approximately forty percent), and YouTube (approximately twenty percent). Less common but present in the cohort were Pinterest, Threads, and X. The cross-platform composition is comparable to the broader Native customer base and is consistent with Northern European SMB social media habits as documented in independent industry surveys.

A.4 Pre-Native baseline distribution

Pre-Native monthly reach baselines varied substantially across the cohort. The median pre-Native monthly reach was approximately 2,200 unique accounts per month. The bottom decile of accounts had pre-Native reach below 100 monthly accounts (including the nine new-activation cases at or near zero). The top decile had pre-Native reach above approximately 25,000 monthly accounts. The wide variance in baseline is what produces the wide variance in lift ratios reported in section 6.5; accounts with very low baselines mechanically produce very high lift ratios when they grow even modestly in absolute terms.

Appendix B: Methodological Notes on Reach Estimation

This appendix provides additional detail on how organic reach was measured for the cohort analysis described in chapter five, and on the limitations of platform-supplied reach metrics that affect the interpretation of the findings.

B.1 Source of reach data

Reach figures for each cohort account were retrieved from the analytics APIs of the platforms on which the account published. Meta’s Graph API supplied reach figures for Instagram and Facebook. LinkedIn’s Marketing Developer Platform supplied reach figures for LinkedIn. TikTok’s Business API and Display API supplied reach figures for TikTok. YouTube Analytics supplied reach figures for YouTube. The data was retrieved on a per-post basis and aggregated to monthly totals using the four-week-month convention described in section 5.4.

B.2 Definition of reach across platforms

The major social platforms define reach differently in ways that are not always fully transparent. Meta defines reach as the number of unique accounts that have seen a piece of content. LinkedIn defines reach similarly but applies different filters around what counts as a “view.” TikTok historically used a more permissive definition that has been tightened in recent platform updates. YouTube’s reach metric counts unique viewers within a configurable time window. The consequence of these differences is that cross-platform aggregation of reach figures may be slightly miscalibrated. We have not attempted to normalise across platforms beyond the four-week-month convention. The internal validity of the within-account before-and-after comparison is unaffected by the cross-platform calibration issue, since the same platform measures the same account using the same method in both periods.

B.3 Treatment of organic versus paid reach

The reach figures reported in this paper are organic reach only. Reach attributable to paid promotion (boosted posts, ad campaigns, sponsored content) is excluded from both the pre-period and post-period measurements. This distinction matters because some cohort businesses also run paid social campaigns, and conflating organic and paid reach would attribute the effects of paid spend to the Native platform. Where platform APIs do not cleanly separate organic and paid reach, we have used the most conservative available filter to ensure that paid reach is excluded.

B.4 Treatment of viral outliers

A small number of cohort accounts had pre-Native or post-Native periods that included a single piece of content with reach more than ten times the median per-post reach for that account. These viral outliers can substantially distort the comparison if not handled carefully. Our default treatment is to include viral outliers in both periods (since they represent genuine reach events for the account in question) but to flag accounts where the inclusion materially affects the lift ratio. In the data reported in chapter six, no cohort account’s headline lift ratio depends materially on viral outlier treatment; sensitivity analyses with outliers excluded produce lift magnitudes within approximately ten percent of the headline figures.

B.5 Engagement attribution

Engagement figures (likes, comments, shares, saves) were retrieved from the same platform APIs as reach figures. Engagement on content that is older than the platform’s analytics retention window cannot be retrieved historically; we have addressed this by restricting the pre-Native window to the most recent ninety days prior to each account’s index date, which is comfortably within all platforms’ retention windows. The consequence is that engagement on very old pre-Native content is not included; if the pre-Native content was producing significant residual engagement from older posts, our measurements understate the pre-Native engagement and overstate the engagement lift. We have spot-checked the magnitude of this potential bias on a sample of accounts and found it to be small (less than five percent of the headline engagement lift).

Native Research, May 2026