Can generative AI make better ad images? What a Journal of Marketing study found
Trained on the best-rated ads in its category, an open-source image model produced ads that beat conventionally made ones with consumers and on Meta. Here is how it worked, how to apply it, and where it fails.
Fine-tuned on top-rated ads, generative AI made ad images that beat conventional ads in pretests and lifted Meta click-through by 49%. How to apply it, and when it fails.
Marketing in 3 · Issue 1 · Week of 12 January 2026
Car advertising was the main test case in the study. Photo: Ariel Acosta on Unsplash.
The short answer
Yes, for mainstream products. When researchers fine-tuned an open-source image model on the 30 best-rated car ads in a set of 543, the ads it produced scored higher with consumers than conventional ads and earned a 49% higher click-through rate than the brand's own ads on Meta. Plain prompting did not get there, and the approach failed for a highly distinctive product.
Key takeaways
Fine-tuned AI ads scored 4.55 out of 7 on attention, interest, desire and action, against 3.79 for conventional car ads; 47 of 50 AI ads beat the conventional average.
On Meta, AI ads drew a 2.04% click-through rate against 1.37% for the brand's real ads (28,775 impressions); on Taboola, 1.41% against 0.64%.
What made the difference was training on the category's best-performing ads. Prompting an off-the-shelf model, or training on random ads, did no better than the brand's own ads.
It worked for an unknown brand and for sunscreen, not only cars. It did not work for a niche, highly differentiated car, and it could not make ads funny.
What the researchers did
Mark Heitmann, Tijmen Jansen, Martin Reisenbichler and David Schweidel asked a question many marketing teams
are asking: can an image model produce ads that work as well as ads made the conventional way? Their answer
comes from seven studies, mostly on cars, with checks on sunscreen and dog care.
They started from 211,429 online ads and found 543 paid car ads. 572 potential car buyers rated these ads on
the four classic funnel stages: attention, interest, desire and action (AIDA), on 1–7 scales. The team then
took the 30 best-rated ads, added about two dozen product photos of the Polestar 3 (a car the model had never
seen), and fine-tuned the open-source model Stable Diffusion on both. Each concept got its own trigger word,
so a single prompt could ask for "a high-AIDA web banner ad of a Polestar 3".
The model then produced 50 ads with no human editing and no prompt crafting. Another 569 car buyers rated
them alongside the competitors' ads and Polestar's own ads, more than 20,000 ratings in all.
Data
543 paid car ads, rated by 572 car buyers
Model
Stable Diffusion 2.1, fine-tuned on the 30 top-rated ads plus 23 product photos
Output
50 AI ads, used as generated
Tests
7 studies; surveys with 127 to 998 people each, plus Meta and Taboola field tests
What they found
The AI ads won on average. They scored 4.55 against 3.79 for competitors' ads and 3.80 for Polestar's own
ads. Only 3 of the 50 AI ads fell below the conventional average, so a randomly picked AI ad was very likely
to be at least as good as a typical human-made one. The best human-made ad (5.57) still beat the average AI
ad, but the best AI ad scored higher still (6.00).
This was not just prettier pictures. The researchers scored every image for aesthetic quality: the most
beautiful conventional ads scored lower on AIDA than the AI ads, and aesthetics explained only about 15% of
the AI advantage. People were also more willing to share the AI ads (4.84 against 3.98 for competitors).
Average rating across attention, interest, desire and action
Average rating across attention, interest, desire and action
Group
Value (out of 7)
Competitors' car ads (543)
3.79
Polestar's own ads (10)
3.8
AI ads, fine-tuned on top ads (50)
4.55
Study 1, 569 potential car buyers, 1–7 scales.
Did it work outside the survey?
Yes. In a four-day A/B test on Meta, five AI ads and five real Polestar ads ran with the same caption, the same
audience (US adults aged 25 to 65 interested in cars) and a $5 daily budget per ad. The AI ads drew more
clicks: 2.04% against 1.37%, a 49% lift. A second test on Taboola showed the same pattern (1.41% against 0.64%).
The authors note that field tests give less control than surveys and that results will vary by platform.
Click-through rate in live tests, real ads vs AI ads
Click-through rate in live tests, real ads vs AI ads
Group
Value (%)
Meta: Polestar's own ads
1.37
Meta: AI ads
2.04
Taboola: Polestar's own ads
0.64
Taboola: AI ads
1.41
Study 3. Meta: 28,775 impressions, 496 clicks, 1–4 February 2025.
How you use the AI matters more than which AI
The most useful finding for practice comes from Study 4. The team compared four ways of using the same model.
Simply prompting it for "an engaging web banner ad" scored 3.99. Training it on randomly chosen ads from the
category scored 3.82, about the market average. Training it on the best ads while also telling it to avoid the
worst ones scored 4.21. Only training on the best-rated ads reached 4.61, the one approach that beat the
brand's own ads (4.35).
In other words, the model has to learn what works with your customers, not just what ads in your category
look like. That takes one round of consumer ratings at the start; after that, new variants cost almost nothing.
Same model, four ways of using it: average funnel rating
Same model, four ways of using it: average funnel rating
Group
Value (out of 7)
Competitors' ads
3.48
Trained on random category ads
3.82
Prompted, no ad training
3.99
Trained on best and worst ads
4.21
Polestar's own ads
4.35
Trained on the best-rated ads
4.61
Study 4, 359 potential car buyers, 1–7 scales.
Brand personality and targeting
The model can also carry a brand image. Adding about 24 Flickr images tagged "rugged" or "luxurious" to the
training produced ads that looked more rugged (4.91 against 4.09 for Polestar's ads) or more luxurious (5.44
against 4.50), with no loss on the funnel measures.
There is a catch for targeting. Rugged-style ads did nothing for the average viewer (4.95 against 4.84), but
they lifted brand liking among people who want a rugged car and lowered it among people who do not. A
personality-tuned version should go to the segment that values that trait, not to everyone.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
Collect your category's ads. Gather a few hundred current paid ads in your category: yours and competitors'. Remove text and logos so the model learns imagery, not words.
Rate them once with real customers. Ask target customers to rate each ad on attention, interest, desire and action (1–7). Five or more ratings per ad were enough in this study.
2
SetBuild it
Train on the top 30 only. Fine-tune on the 30 best-rated ads, not on a random sample and not on a mix of best and worst. This was the decisive step.
Teach it your product and brand. Add 11 to 24 clean product photos from several angles, background removed, so the product is drawn accurately. For a personality such as rugged or luxurious, add about two dozen images that express it, under its own trigger word.
3
MarketLaunch and measure
Generate many, then screen. Produce dozens of variants and screen out distortions, odd details and anything off-brand before anyone sees them.
Test live before scaling. A/B test the best AI images against your current creative on the platforms you use, and scale only what wins on click-through.
What to expect
A better average, not a guaranteed winner. AI ads beat the typical conventional ad by about 0.75 points on a 7-point scale; your best human ads can still win.
Lower cost per variant. The authors cite $2,000 to $4,000 for a simple stock-based ad and up to $50,000 for a professional shoot. After the one-off rating and training, extra variants cost little, which helps fight ad fatigue.
A click-through lift worth testing, not banking. The 49% lift came from a small four-day test with one car. Treat it as a reason to run your own A/B test.
Stable settings. The same training settings worked for cars and sunscreen, so the process can be standardised once set up.
Where it does not work
Distinctive products. For the Smart Fortwo, a niche two-seater sold on surprise and city agility, AI ads scored slightly below the real ads (3.62 against 3.81). The model learns the category average; it cannot invent an unusual angle.
Humour and surprise. Trained on funny dog-care ads, the model kept good funnel scores but could not make ads funny. Pedigree's real humorous ad was recognised more often than its AI version (85% against 70%).
Sameness. AI ads in a category looked alike: similar roads for cars, the same Monstera leaves in sunscreen ads. If every competitor trains on the same winners, ads may converge and lose impact.
Fit with the goal. The model improves what it is trained on. It did not raise recall or recognition beyond conventional ads, because it was never trained on them.
Questions marketers ask
Can generative AI make ads that perform better than human-made ads?
In this Journal of Marketing study, yes on average: fine-tuned AI ads scored 4.55 out of 7 on attention, interest, desire and action, against 3.79 for conventional car ads, and earned a 49% higher click-through rate on Meta than the brand's real ads. The best human-made ads can still beat the average AI ad.
Is prompting an image generator enough to get good ads?
Not in this study. Prompting Stable Diffusion for an engaging banner ad scored 3.99, below the brand's own ads (4.35). Only fine-tuning the model on the 30 best-rated ads in the category beat them (4.61).
How much data do you need to fine-tune an image model for ads?
The researchers used the 30 top-rated ads from a set of 543 rated by consumers, 11 to 24 product photos, and about 24 images per brand personality trait. The consumer ratings are collected once.
When does AI-generated advertising not work?
It worked less well for a highly differentiated niche product, could not create humour, and tended to produce similar-looking ads. It suits mainstream products and goals built on familiarity and trust.
Does it only work for well-known brands?
No. AI ads for NIO, a brand unknown to the study's participants, scored 4.94 against 4.02 for NIO's real ads, and AI sunscreen ads beat both competitors and the brand's own ads.
The counterpoint to this week's lead: in three experiments, women rated a model as less attractive, and liked the ad less, when a disclaimer said the model was AI-generated rather than real. The label triggered resistance and made the model feel more distant, for Gen Z and Gen X women alike. If you use AI-generated people and must disclose it, test how your audience reacts to the label itself.
Chen, Q., Greene, D., Liang, Y., Portmann, M., & Dolnicar, S.
In an online experiment with 995 people, images made by AI models were rated more accurate, more appealing and more emotionally clear than images made by human designers, and people struggled to tell them apart. ChatGPT-4o was best at consistent sets in a simple style; complex emotions such as envy were still hard to draw. Useful for teams that need many controlled visual variants for testing.
Source
Heitmann, M., Jansen, T. P. J., Reisenbichler, M., & Schweidel, D. A. (2026). Picture perfect: Engaging customers with visual generative AI. Journal of Marketing, 90(4), 74–96. https://doi.org/10.1177/00222429251356993
This brief is written for practitioners from the published article. Open access (CC BY 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
Placing ads by both look and topic of the page lifted predicted attention by up to 10%; matching by category or text alone did not. Evidence from 3,531 ads.
Viewers who worked out the actor was AI rated the ad 2.83 against 4.59. Being told afterwards hurt far less, and priming people made them twice as likely to notice.
A plain "this ad is deepfaked" label changed nothing. Adding one sentence about what that means cut perceived truthfulness from 3.33 to 2.41, without costing a single sale.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
