Can background images studied for one company’s Instagram campaign score highly for both expected clicks and brand fit? Only 5.3% of the images used to train the system ranked in the highest-scoring quarter on both measures. For a business, a high score for expected clicks does not necessarily mean an image meets its brand requirements.
ACADEMIC WORK REVIEWED
Leveraging Generative AI to Create Visual Content in Digital Advertising
ResearchersRemi Daviet, Yohei Nishimura
Marketing Science / SSRN · 2 April 2026 · Examined SSRN working version; journal issue September 2026 · Peer-reviewed article · Source language: English
English review: 20 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?
Can AI select advertising backgrounds that meet brand requirements while also offering strong click potential?
The researchers address the selection problem that follows large-scale image generation. Their aim is to find brand-acceptable options while learning from costly field tests to predict link click-through rates. The study asks whether visual appeal is sufficient to identify suitable, effective advertising images.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Visual appeal alone may give an advertising team too little information to make a selection. Keeping brand approval criteria separate from click performance makes the reason for accepting, rejecting or selecting an image easier to explain.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Remi Daviet and Yohei Nishimura ran an Instagram campaign with an outdoor activities company whose name and country are not explicitly stated. Training took place in Q4 2023. The audience comprised residents aged 18 to 55 with children aged 3 to 12. Placements were restricted to the feed, Explore and search results. Initial brand training used 60 batches of 12 images, while performance training involved 9 trials, each with 12 images and controls. Ultimately, AcceptAI was trained on 1,417 images and PerfAI on 125. The final comparison contained 10 images in each group: selections from the system, one professional designer and a model optimising only aesthetic scores. The groups ran during the same week with identical budgets and targeting. This was not a conventional randomly assigned A/B experiment.
Understanding the method · DijitalPi commentary
A custom landscape generator was combined with SDXL to improve image quality. AcceptAI predicted whether the brand would accept an image, while PerfAI predicted its link click-through rate. A selection process called active learning chose images expected to provide useful training information in each round. Immediate click maximisation was not the training objective. The company required landscapes that look like photographs representing its service areas without AI-generated defects. Initial brand-training labels took the company representative less than two hours to provide.
All images in the final comparison received company approval. The logo and copy stayed fixed, with only the background changing. The aesthetics group used the same generator, guided by an aesthetic score from TANet. The primary measure was link click-through rate: link clicks relative to advertising impressions. Sales and bookings were not measured. The researchers treated the platform’s delivery algorithm as part of the environment, evaluating advertising performance within that system rather than isolating a causal effect of the image alone.
03 / RESEARCH FINDINGS
Brand acceptability and click potential showed limited overlap
The researchers found that 5.3% of training images ranked in the top quarter for both performance and brand acceptability. The overlap in the top tenth for both measures was 0.6%. In the final comparison, mean link click-through rates were 0.98% for the system, 0.78% for the aesthetics group and 0.65% for the designer. The system’s comparison with the designer was statistically significant (p = 0.0041). Comparisons between the system and aesthetics group (p = 0.0924), and between the aesthetics group and designer (p = 0.1971), did not meet the 5% significance threshold. Results varied less within the system’s group, which the authors interpreted as reduced creative risk. In April 2024, without retraining, the system recorded a mean of 3.38% against 3.24% for the company team’s campaign. No significant difference was detected (one-tailed p = 0.323). The company’s highest-performing individual advertisement achieved 5.39%, exceeding the system’s highest result.
Images in the top quarter for both measures
5.3%The share of training images ranking in the top quarter for both performance and brand acceptability. This is not the overall brand approval rate.
System group’s mean link click-through rate
0.98%The average across 10 images in the final comparison, not a sales or booking rate or an expected business gain.
DijitalPi commentary: A practical application is to manage image selection through two separate questions: does the business approve this image, and how do approved options perform under the current advertising setup? Document approval criteria, exclude unsuitable images and assess the group average separately from a single attention-grabbing advertisement. This is a proposed workflow to test before considering implementation of the paper’s custom system.
One company, one campaign, one platform and one designer comparison do not support a broad claim of superiority. Identical budgets and targeting do not remove the influence of the platform distributing images differently. The absence of significant differences against the aesthetics model and in April 2024 also does not establish equivalence. The same performance cannot be assumed for Türkiye, Turkish-language advertisements, product images or complex human figures; implementing the custom system also requires substantial technical expertise.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The country where the study took place is not specified. The measured outcome was link click-through rate, not sales or bookings. The study suggests that a system combining brand acceptance with link-click prediction can produce usable results in the examined campaign. The significant comparison with the designer’s group cannot be generalised to all designers or sectors, and sales effects were not measured. The article was published in the peer-reviewed journal Marketing Science, but this review uses the SSRN working version dated 2 April 2026, which may differ from the journal text. The authors are academics at the University of Wisconsin-Madison, and the field study involved an unnamed company. No funding or competing-interest statement was found in the examined text. This does not establish that funding or competing interests were absent.
Our open question at DijitalPi: which image requirements would you refuse to compromise for clicks in a Turkish-language campaign, and which separate business outcome would you use to assess approved backgrounds?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
Tourism: the landscape does not match the itinerary
An illustrative tour operator is choosing an advertising background from AI-generated landscapes.
Does the attention-grabbing image accurately represent the place customers will visit?
Open the recommendation for example 01+
Check the service area first, then compare approved images.
Check each background against the actual itinerary and the business’s visual requirements. Keep the logo and copy fixed when measuring link click-through rates for approved options. Record bookings as a separate outcome. The paper’s click results do not establish that this business will gain bookings.
EXAMPLE 02
Local services: an impressive garden creates the wrong expectation
An illustrative garden maintenance business has advertising backgrounds showing unsuitable local plants and landscaping services it does not offer.
Could the image attract interest while misrepresenting our service?
Open the recommendation for example 02+
Include local suitability and service scope in image approval criteria.
Have the business owner assess each image for local suitability, service scope and AI-generated defects. Exclude rejected images from the comparison even if their predicted click potential is high. Test approved backgrounds under consistent advertising conditions and track link clicks separately from quotation requests. The research did not measure effects on enquiries for local services.
EXAMPLE 03
B2B software: the background changes the apparent product promise
An illustrative B2B software company is generating backgrounds for its advertisements.
Could an eye-catching image suggest features that the software does not have?
Open the recommendation for example 03+
Keep the verified product screen intact and limit the test to background selection.
Hold the actual product screen and advertising copy constant. Check whether any background implies an additional feature or an inaccurate product promise. Measure link click-through rates for approved options while recording demo requests separately. This study did not test product interface generation or B2B customer acquisition.
TRY IT WITH YOUR TEAM
Write your image approval criteria
Plan a small comparison that assesses brand acceptability and link click-through rates separately, changing only the background. Keep targeting, bidding settings, budgets and placements constant, and account for the platform’s delivery system when interpreting results. This is an illustrative test proposal that DijitalPi has not implemented.
SMMACHINE creates visuals and copy to match your brand’s voice and checks content quality. These checks do not show how often an image will be clicked. Explore SMMACHINE ↗
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