In Taboola’s advertising data, did artificial-looking ad images receive fewer clicks? Images perceived as more artificial had lower click-through rates, and this relationship was stronger for AI-generated images. The initial gap between click-through rates of 0.76% for AI images and 0.65% for human-made images disappeared when similar ads were compared. For businesses, this suggests that whether an image looks artificial to viewers matters more than who produced it.
ACADEMIC WORK REVIEWED
AI in Disguise—Quasi-Experimental Analysis of a Large-Scale Deployment of AI-Generated Ads
English review: 19 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Among comparable advertisements, does click-through rate vary with AI image production and how artificial the image looks?
The researchers compared images produced with Taboola’s GenAI Ad Maker against comparable human-made advertisements. They sought to inform decisions about using the tool and identify visual features associated with an image being perceived as AI-generated.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Choosing a production method and selecting an image for publication are different decisions. This study suggests that production origin alone may be an insufficient selection criterion. Assessing whether an image looks artificial provides an additional question to investigate when interpreting click data, without determining the business’s sales outcome in advance.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The researchers examined English advertisements distributed across publisher websites by US-based Taboola from 3 June to 15 August 2023. The main comparison covered 4,633 advertisements in 1,186 quasi-experiments across 351 campaigns run by 202 advertisers, comprising 29,631 ad-day observations, more than 369 million impressions and 2.5 million clicks. Here, a quasi-experiment means comparing similar advertisements without random assignment. Each of 1,751 unique images was also assessed for perceived artificiality by 5 US-based MTurk participants, with 612 different raters participating overall.
Understanding the method · DijitalPi commentary
The researchers define sibling advertisements as variants launched by the same advertiser, in the same campaign, on the same day. Groups containing both AI-generated and human-made images share the advertiser, campaign, landing page, objective and launch date. Each comparison ends when the first advertisement in its group stops running. The analysis also accounts for date, copy and image differences. The primary outcome is click-through rate: clicks divided by impressions. The full dataset of 305,121 advertisements and more than 16.4 billion impressions should not be confused with the main comparison sample.
GenAI Ad Maker became available to all English advertisers on 12 July 2023, using Stable Diffusion 2 to generate images. Each request produced 4 candidate images. The advertiser chose which one to publish. Researchers could see only that selection. Viewers were not told that images were AI-generated. For the artificiality assessment, participants rated whether an image was human-made or AI-generated on a 5-point scale. This did not measure purchase intention.
03 / RESEARCH FINDINGS
Perceived artificiality helps explain click differences beyond production origin
Before accounting for differences between ads, the click-through rate in the main sample was 0.76% for AI-generated images and 0.65% for human-made images. This raw advantage disappeared in the sibling-ad comparison: the average difference for AI images was small, negative and statistically insignificant. This establishes neither average superiority nor disadvantage nor equivalence. Greater perceived artificiality was associated with lower click-through rates, with a stronger relationship for AI-generated images. The uncertainty interval for the separate human-made estimate included zero, so the same relationship cannot be described as clearly established within that group. In the model, AI images perceived as less artificial were associated with higher click-through rates than their human-made siblings. This advantage diminished and reversed as perceived artificiality increased. A supplementary analysis measured conversion rate as conversions divided by impressions: the separate main effects of image origin and artificiality were insignificant, while the interaction representing their joint relationship was negative and significant.
Raw click-through rate for AI-generated images
0.76%The uncontrolled rate in the main sibling-ad sample. The apparent advantage does not remain after comparability controls.
Raw click-through rate for human-made images
0.65%The uncontrolled comparison value in the same sample, not a sales or conversion rate.
DijitalPi commentary: Ask about perceived artificiality alongside liking when reviewing images before publication. The researchers associated greater aesthetics and colour saturation with perceived AI production, and realism, clarity and warmth with perceived human production. Turn these features into questions to test on your own images rather than automatic editing rules. Keep perception assessments separate from advertising performance, and clicks separate from realised business outcomes.
These associations do not establish that reducing colour saturation or adding a face will increase clicks. Feature relationships may depend on the category. Advertisers chose both whether to use the tool and which candidate image to publish, narrowing the findings’ scope. Stopping the first advertisement because of early performance could influence the comparison window. The authors report that checks addressing this did not change the results. Images generated with other AI tools and uploaded manually also could not be identified.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The study observed existing ads rather than randomly assigning them to groups. This SSRN working paper finds no detectable average click-through disadvantage for AI images among comparable advertisements and supports examining perceived artificiality separately. There is no information establishing it as a peer-reviewed journal publication. Its findings concern one platform, Stable Diffusion 2, the 2023 launch period, advertisers who chose the tool and English advertisements without disclosure of production origin; they cannot be assumed to apply to Turkish advertising or Türkiye. Clicks may reflect curiosity or attention and do not establish increased sales. The authors disclose collaboration with Taboola to obtain data, no compensation from the company, no funding to report and no financial conflicts of interest. The paper presents its findings within a framework supporting Taboola’s launch communications and advertiser guidance.
Our open question at DijitalPi: if perceived artificiality is associated with click-through rate in your Turkish advertisements, does that relationship also appear in qualified enquiries or purchases?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: oversaturated colours in a product image
An illustrative furniture retailer is assessing advertisements showing the same sofa with natural and oversaturated colours.
Are striking colours associated with the image looking artificial?
Open the recommendation for example 01+
Check product colour accuracy and assess perceived artificiality separately.
Check that each image represents the actual product colour accurately and ask the target audience how artificial it looks. For the advertising comparison, hold copy, landing page and delivery period constant, recording click-through rate and purchases separately. The study’s association with colour saturation does not establish that reducing saturation will bring this retailer more clicks or sales.
EXAMPLE 02
Tourism: an accommodation advertisement looks too perfect
An illustrative accommodation business is choosing between a photograph and an AI-generated advertising image of the same room.
How can visual appeal be distinguished from perceived artificiality?
Open the recommendation for example 02+
Ask separate questions about liking and artificiality.
First check that both images accurately represent the room’s actual features. Ask the target audience to assess liking and artificiality separately, then track click-through rate and completed bookings independently during the campaign. The paper’s association with aesthetics does not mean every attractive advertisement will receive fewer clicks.
EXAMPLE 03
B2B software: image selection and enquiry quality
An illustrative business software provider plans to publish the same advertising copy with human-made and AI-generated images.
Is a more natural-looking image also associated with qualified enquiries?
Open the recommendation for example 03+
Treat clicks and enquiry quality as separate success measures.
Assess perceived artificiality before publication. When comparing advertisements leading to the same landing page, record click-through rate, enquiry numbers and whether enquiries meet predefined suitability criteria. Since the study’s primary outcome was clicks, it cannot establish gains in qualified enquiries or signed contracts.
TRY IT WITH YOUR TEAM
Plan your image comparison
Prepare an artificiality assessment for images intended for the same campaign, keeping copy, landing page and delivery conditions consistent in the comparison. Plan to measure the business outcome that follows a click as well as the click-through rate. This is a proposed pilot that DijitalPi has not conducted.
You can also apply this review to your social media visuals. SMMACHINE prepares visuals for your brand and supports readability and layout checks. Measure which visual attracts more clicks in your own campaign. Explore SMMACHINE ↗
Yannick Exner, Jochen Hartmann, Ziqian Ding, Shunyuan Zhang, Oded Netzer TUM School of Management (Exner, Hartmann); Carnegie Mellon University (Ding); Harvard Business School (Zhang); Columbia Business School (Netzer).
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences