Imagine you run marketing for a car brand. You want a striking online ad, so you hire a photographer, book a location, license some imagery, and run focus groups to see if the result resonates. The bill can climb past $50,000 for a single shoot. Now imagine you could feed consumer reactions directly into a computer and produce an endless supply of ads that perform at least as well, at a tiny fraction of the cost.
A team of marketing researchers set out to test whether that scenario holds up. Writing in the Journal of Marketing, Mark Heitmann and Tijmen Jansen of the University of Hamburg, Martin Reisenbichler of the Vienna University of Economics and Business, and David Schweidel of Emory University describe a series of experiments designed to see whether generative artificial intelligence can create visual advertising that connects with consumers.
The gap between clever software and marketing goals
Image generators like Stable Diffusion can produce convincing pictures from a text prompt. Ask for “a dog sitting on a car hood,” and the software will iteratively transform random noise into a coherent image. These tools are trained on enormous collections of labeled pictures, so they know what a dog or a car looks like.
What they do not know is what makes an ad effective. As the researchers point out, an off-the-shelf model was never designed with a marketer’s objectives in mind. When one Google model was asked to “generate an effective advertising image,” it responded that doing so would be “impossible because it’s a subjective concept that wouldn’t translate well into a visual representation.”
The authors wanted to close that gap. Advertisers often measure success using what are called mindset metrics, which capture stages a consumer moves through. A common framework is AIDA, short for attention, interest, desire, and activation. The team asked whether an AI model could be trained not just to draw a car, but to draw a car ad that scores well on those four dimensions.
Teaching the model what works
Their approach relied on a technique called fine-tuning, which means taking a general-purpose model and giving it extra, targeted training. The workflow had a few steps. First, the researchers gathered images: clean product shots of the vehicle they wanted to feature, plus a large collection of real online car ads.
Next, they measured how consumers responded. They collected online banner ads, stripped out the text to avoid glitches, and had survey participants rate them on seven-point scales for each AIDA phase. Attention was measured with a statement like “This advertisement would stand out in comparison to other advertisements,” and activation with “If I were in the market for a car right now, I would consider buying the car in this advertisement.”
They then took the highest-rated ads and used them to fine-tune Stable Diffusion, linking the visual patterns of successful ads to a unique text label the model could later be prompted with. The idea was to let the software learn the visual language of high-performing ads and reproduce it around a new product, without any human doing creative prompting or editing.
For their main tests they chose the automotive industry, and specifically the Polestar 3, a vehicle recent enough that it did not appear in the standard model’s training data. That meant the AI could not simply copy an existing ad. The team collected more than 211,000 online ads to build their material and recruited hundreds of licensed drivers through the Prolific research platform to rate the results.
What the ads scored
Across the first several studies, the AI-generated ads tended to outperform conventionally produced ones on the mindset metrics. In the opening study, real online car ads averaged 3.79 out of seven, while the AI-generated ads averaged 4.55. Of the 50 AI ads, 47 scored above the average conventional ad. The single best human-made ad still beat the average AI ad, which the authors interpret as a sign that human creativity can produce standout work, but the best AI ad edged it out.
The team also checked whether the results were simply a matter of prettier pictures. Using an automated aesthetic-quality score, they found that the most visually attractive conventional ads actually scored lower on AIDA than the AI ads. The authors interpret this as evidence that attractiveness alone does not explain the gap; training on effective ads appears to add something beyond good looks.
A second set of studies tackled brand image. Advertisers often want to convey a specific personality, and the researchers tested two opposing ones: “ruggedness” and “luxury.” By adding a small set of images tagged with each trait, they could steer the generated ads toward the desired feel. Ads trained on ruggedness were rated as more rugged, ads trained on luxury as more luxurious, and, importantly, adding this personality training did not drag down AIDA performance. The team also found they could tailor ads to consumer segments: people who said they valued ruggedness responded more favorably to the rugged-styled ads.
To move beyond survey ratings, the researchers ran the ads as real campaigns on Meta and on another platform, Taboola. Here the measure was click-through rate, the share of people who see an ad and click it. Across roughly 28,000 impressions on Meta, the AI-generated ads drew a click-through rate of 2.04% versus 1.37% for the real Polestar ads, an increase of nearly half. The pattern favored the AI ads on the second platform too, though the authors caution that click rates vary and should be read alongside other measures.
What matters, and where it breaks down
One study compared shortcuts to the full workflow. Simply prompting the standard model, or fine-tuning it on randomly selected ads rather than top performers, produced weaker results. The authors interpret this as evidence that consumer feedback is the key ingredient. It is not enough to imitate the general look of car ads; the model needs to learn the look of the ads that actually work.
The team also probed the limits. Brand familiarity did not seem to matter: the approach worked for an unknown Chinese electric-vehicle brand just as well as for Polestar. It also worked for a fast-moving consumer product, sunscreen, which appears in a wider variety of settings than cars.
The clearest boundary showed up with the Smart Fortwo, a tiny two-seater whose ads often rely on surprise to highlight its unusual size. For that highly differentiated product, the AI ads scored slightly lower than the real ones. The researchers explain that their method essentially learns a visual average of what works in a category, which builds a sense of fluency, familiarity, and trust. Follow-up tests showed AI ads were rated as easier to process, and this fluency correlated with performance. But averages, by design, do not produce the incongruity that humor or unexpected creative ideas depend on. When the team tried to train the model on humorous dog-food ads, it could not reach the humor levels of the human-made originals.
What it might mean for marketers
The authors suggest that parts of the advertising process could shift. Because consumer feedback needs to be gathered only once to fine-tune a model, marketers could then generate a large number of ad variants aligned with their goals, personalize them for different audiences, and refresh them often to fight the fatigue that sets in when people see the same ad repeatedly.
The researchers also flag a risk. As more advertisers use similar tools trained toward similar objectives, they noticed recurring motifs, such as Monstera leaves appearing across sunscreen ads and near-identical road settings across car ads. If everyone’s AI converges on the same visual averages, they argue, ads may grow more homogeneous and, over time, less effective. The takeaway they offer is one of balance: generative AI can handle a lot of the routine execution, but distinctive products and bold creative messages still tend to call for a human touch.
