What does the evidence establish about AI-generated advertising effectiveness?

Production methods, audience response and the distinction between preference and commercial increment

This retrospective review considers source versions available by 20 October 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: AI advertising; creative effectiveness; persuasion; incrementality

1. Research question and scope

The question of whether AI-generated advertising works contains several different comparisons. A model can generate an appealing image, help a professional finish a film, or replace part of an existing advertising workflow. Those interventions have different costs and may produce different outcomes. A comparison between a synthetic banner and stock photography does not estimate the difference between an autonomous campaign and a campaign developed by a full creative team. The treatment needs a production description before its result has an interpretable scope.

This review asks what the selected empirical literature establishes about audience response to AI-generated advertising and related persuasive communication. It distinguishes evaluations of the material, behaviour during advertising exposure and incremental commercial effects. It also separates actual content origin from viewers’ beliefs about origin. A person can prefer a generated ad without identifying it as generated; conversely, a disclosure can alter evaluation while the image and copy stay identical. Those are different experimental questions [1] [2] [12].

The relevant standard of comparison is contextual. “Human-made” may refer to a freelancer responding to a brief, a stock photograph, a lay participant composing a message or a professionally produced campaign. The expression supplies no common quality threshold. Likewise, a generated asset can be selected from many candidates, revised by people or accepted on the first attempt. These choices affect what the comparison represents and how production efficiency should be interpreted.

2. Review method

Targeted web searches on 21 October 2026 used the terms AI-generated advertising, visual marketing, field experiment, persuasion, disclosure and advertising measurement, followed by searches for named studies. The review consulted publisher abstracts, author-institution records and accessible manuscripts. 17 sources were selected for their relevance to direct advertising outcomes, adjacent persuasive tasks or the interpretation of measurement. This is a focused narrative review; it does not claim exhaustive coverage, independent duplicate screening or a preregistered synthesis protocol.

The extraction distinguishes publication versions. The current Columbia record for Exner and colleagues is dated 20 October 2026 and describes a substantially larger observational setting than an earlier sibling-ad report circulated under a related title. The current record is used here. Numbers from the two versions are not combined. Where a source was accessible only as an abstract or metadata record, the review confines itself to the reported design and outcome; it does not infer unexamined robustness checks [2].

No common effect size is calculated. Preference shares, click-through rates, attitude changes and rated aesthetic quality have different denominators and meanings. The evidence map records the outcome directly studied and its principal boundary. It does not rank studies by the size of a headline percentage. The methodological advertising papers are included to explain identification and uncertainty, rather than to supply an estimate of the effect of generative production.

3. Visual quality and selected advertising comparisons

Hartmann, Exner and Domdey study visual marketing content produced by several text-to-image systems. Their published abstract describes 10,320 synthetic images, 2,400 human-made input images and 254,400 human evaluations. The generated material could exceed human-made images on assessed quality, realism and aesthetics; the research also includes a banner-ad application. These results establish that synthetic provenance does not prevent strong performance on those measures. The comparison is bounded by the specified systems, assets and evaluation procedure [1].

The count of evaluations should not be treated as the count of independently sampled campaigns. A large number of ratings can provide precision about a fixed stimulus collection while leaving uncertainty about other collections. An evaluation of selected images also leaves open how often the generator produces an acceptable candidate and how much human selection was required. Both questions matter when moving from a stimulus experiment to a production process.

The title’s reference to superhuman content requires care. A generated image can outperform the particular human comparator without outperforming the best available commissioned work. A stock image answers a general market need; a brand-specific photograph can answer a different brief. Matching the production objective, budget and selection opportunity would be necessary for a broader comparative claim. The abstract-level material consulted here does not establish such a universal ranking.

Visual realism is also a limited criterion for an advertisement. A realistic product image can still communicate the wrong feature or obscure the advertiser. Conversely, an intentionally stylised image may suit the creative task. The useful empirical question concerns the prespecified response in the intended audience. Treating one aesthetic score as an overall measure of advertising value would erase the distinctions that make the study informative.

4. Evidence from live display advertising

The 20 October 2026 author-institution record for Exner and colleagues describes more than two million ad-day observations, over seven thousand advertisers and nearly fifty product categories, spanning more than sixteen billion impressions and 116 million clicks. The working-paper abstract reports that ads containing generated images achieved higher click-through rates when those images did not look AI-generated. It identifies visual features associated with viewers’ origin perceptions [2].

This source supplies live-platform evidence and a boundary condition concerning perceived origin. The abstract does not establish that every generated image performs better, or that disclosing provenance would leave performance unchanged. The finding about images that do not look generated concerns perception of the asset; a disclosure is an additional intervention. Nor does a click-through advantage automatically identify additional purchases. The current source record foregrounds clicks, so this review does not attach an unreported revenue effect to it.

An observational platform setting contains selection at several levels. Advertisers choose tools, campaigns compete for delivery, and optimisation systems may allocate impressions in response to early performance. Design controls can reduce those concerns, but the abstract alone cannot provide a complete appraisal of identification. The large data volume describes coverage; it does not by itself remove confounding. Causal interpretation depends on the comparison procedure, assumptions and diagnostics in the working paper.

The visual-cue result also warns against treating perceived AI origin as a direct measure of actual production. Viewers infer origin from features that can appear in both generated and human-created images. Conditioning on that inference may divide content by visual style as well as by provenance. A result within the apparently human subset therefore cannot isolate the effect of concealment or prove that an undisclosed generated image would always outperform a labelled version.

5. Preference experiments and persuasive language

Meguellati and colleagues’ preprint reports two studies comparing generated and human-created advertisements. The first concerns personality-oriented advertising; the second compares appeals based on persuasion principles. The author manuscript reports a preference advantage for generated material in the latter experiment. The outcome is a choice between ads under the study procedure, rather than an observed purchase generated by a paid media campaign [3].

The first comparison’s absence of a statistically significant difference should not be called demonstrated equivalence. That conclusion would require an equivalence margin and an appropriate analysis. More generally, preference can reveal comparative response while remaining an imperfect proxy for incremental behaviour. A respondent asked to choose which of two ads is better has attention, a decision prompt and alternatives that may differ from ordinary exposure. None of those features makes the experiment uninformative; they define its estimand.

The relationship between persuasive writing and commercial effect is equally conditional. Bai and colleagues report three preregistered policy-message experiments involving 4,829 participants. Generated messages changed policy attitudes relative to neutral messages and performed similarly to lay-human messages. The study concerns political attitudes and particular text-generation procedures, not branded advertising effectiveness. Its relevance is evidence that generated language can influence reported opinions under controlled exposure [6].

Salvi and colleagues examine GPT-4 in short structured debates. The design compares conversational persuasion and includes personalisation conditions. Its author interpretation describes a proof of concept under controlled conditions. Conversation allows adaptation to a participant’s replies, while a conventional display ad has no such exchange. Transferring the result to a banner, pre-roll video or retail shelf would require evidence about the altered attention and interaction conditions [5].

Persuasion itself needs an outcome definition. A message can change a stated opinion immediately yet leave behaviour unchanged, or affect behaviour without producing a large reported attitude change. Duration matters as well. A post-exposure measurement cannot establish persistence over weeks unless follow-up observations are collected. The reviewed studies support the capability to persuade in specified settings; they leave the durability and commercial consequences of many advertising applications unresolved.

6. Personalisation, provenance and production conditions

Matz and colleagues’ earlier field studies examine psychological targeting in digital mass persuasion. Their results concern matching persuasive appeals to psychological characteristics, with behavioural advertising outcomes. They predate current generative systems. They therefore provide evidence about message-audience fit, rather than a direct test of AI production. A later generator could create both matched and unmatched messages, so personalisation and production technology need separate assignment or analysis [4].

Audience fit also complicates an aggregate creative comparison. A generated ad might be effective for one audience and ineffective for another. Averaging across people can hide those differences; dividing the audience into many small groups can create unstable estimates. A credible heterogeneity claim needs adequate observations, prespecified subgroup definitions or clear separation between exploratory discovery and confirmation. The fact that a system can generate many variants supplies no evidence that it can select the best one for each person.

Actual provenance, perceived provenance and disclosed provenance should be measured separately. A content-origin experiment changes how the material was produced. An origin-belief experiment changes what viewers are told. A perception study observes or measures what viewers infer. Koning and Voorveld examine disclosure and trust as a distinct research problem; the publisher record reports pathways involving persuasion knowledge [12]. That record is used only to establish the relevance of the distinction, without importing a general numerical disclosure penalty.

Human involvement is another treatment component. Prompt formulation, candidate selection, factual checking and editing can affect the final asset. When a study evaluates only the selected output, it cannot estimate the isolated contribution of the model without a design that varies those components. The language of fully automated production should be reserved for procedures that actually remove human direction at the relevant stages. Otherwise the tested unit is a human-system workflow.

7. Why clicks and conversions do not settle incrementality

A campaign can record conversions from people who would have purchased without seeing the ad. Observational attribution attaches a response to an exposure or click according to a rule; causal incrementality compares what happened with what would have happened under a defined alternative. The two quantities can differ substantially. Gordon and colleagues compare observational advertising methods against large Facebook field experiments, showing why access to rich behavioural data does not automatically reproduce experimental estimates [7].

The same issue applies when comparing generated and conventional ads. If a platform shows one variant to people with greater purchase propensity, the observed conversion rate combines creative response with audience selection. A randomised assignment can identify an effect of the assigned policy, while imperfect delivery may require distinguishing assignment from actual exposure. The estimand must name whether it concerns offering the creative, displaying it, or reaching a specified audience under the platform’s delivery system.

Lewis and Rao examine the difficulty of estimating advertising returns precisely. Their analysis concerns the relationship between modest advertising effects and noisy business outcomes. It gives a reason to report uncertainty even when a campaign has a large audience [8]. An imprecise estimate spanning meaningful benefit and meaningful harm does not support a confident claim that the two production methods perform the same. Power depends on outcome variance and the effect of interest, not simply the number of assets generated.

Blake, Nosko and Tadelis’ paid-search field experiment illustrates the importance of consumer heterogeneity and a credible counterfactual in advertising measurement [9]. This is earlier, non-generative advertising research. Its relevance is methodological: observed response among customers already inclined to buy can overstate the effect of the advertising intervention. AI changes the production method; it does not remove the counterfactual problem.

A conversion advantage also needs an economic denominator. More purchases at a higher total acquisition cost can have a different value from the same purchases at lower cost. The relevant cost includes production, media, revision and approval within the defined comparison. Without common accounting boundaries, a reported production saving and a reported response increase cannot be combined into a valid return-on-investment estimate.

8. Adjacent creativity and productivity findings

Evidence about professional writing and short fiction can explain why generative tools are attractive without resolving advertising effectiveness. Noy and Zhang study professional writing tasks and report improvements in speed and assessed quality. Those endpoints describe production performance in the tested work. They do not measure whether consumers recognise a brand or buy more after exposure to an advertisement [11].

Doshi and Hauser examine constrained story writing with access to generated ideas. They find changes in individual creative evaluations alongside increased similarity across stories. Their setting raises a separate question about collective diversity, but it does not establish that greater similarity lowers campaign returns. Advertising may require consistency within a campaign and differentiation from competitors; a story-similarity metric does not measure that balance [10].

These adjacent studies are most useful when their outcome labels are preserved. Faster writing is a productivity result. Higher ratings are an evaluation result. Shared semantic features are a diversity result. Their commercial significance depends on additional evidence about the audience and assignment. Combining them into a single claim that AI makes advertising better would go beyond every individual study’s design.

Bleier and Eisenbeiss examine when and where personalised advertising is effective; generated provenance is only one feature of an exposure. [13] Aguirre and colleagues connect information collection and trust to personalised-advertising effectiveness, making data collection part of the intervention rather than a neutral input. [14] Lambrecht and Tucker distinguish information specificity in retargeting, a separate comparison from whether a model wrote the message. [15]

Rubin’s potential-outcomes formulation places the causal question on outcomes under alternative treatments for a defined population. [16] Ghost-ad experiments address the counterfactual exposure problem in online advertising; they illustrate why creative preference alone cannot identify incremental campaign effects. [17]

9. Synthesis and limitations

The selected evidence establishes that generated marketing content can receive strong evaluations and that generated images can perform competitively in live advertising settings. It also establishes that language models can produce persuasive text under controlled exposure. These are substantive findings. Their scope depends on the asset, comparator, generation procedure, audience and outcome, with no common estimate spanning all those conditions.

The review is selective and several sources were accessible only at abstract or publication-record level. It has not independently reproduced analyses, reconciled every manuscript version or assessed supplementary robustness checks. Publication and platform selection may favour settings in which data are available or tools have already been adopted. A working paper’s scale should therefore be read alongside its status and identification strategy.

The answer to the research question is conditional: AI-generated advertising can work on the outcomes directly tested, but production provenance alone predicts neither success nor failure. Aesthetic quality, preference, click-through response and incremental purchasing remain distinct empirical endpoints. The unresolved question for a particular campaign is the effect of its specified production and delivery policy relative to a credible alternative, measured on the outcome the campaign is intended to change.

Source access and appraisal register

The additional comparative sources in references 13–17 are used for the specific distinctions cited in the discussion. Access is limited to previously inspected primary-source abstracts, publication records or manuscripts. Their inclusion does not imply a fresh full-text risk-of-bias appraisal or direct validation in the application discussed here.

The register describes the material consulted for this narrative review. Access to a manuscript does not imply replication, independent extraction or a formal risk-of-bias assessment. Source dates identify the underlying research.

References