Is Your AI Ad Image Costing You Clicks?

Picture yourself as an advertiser building an online ad campaign on Taboola, a US-based global advertising platform that reaches about 500 million daily active users. The headline is ready, the landing page is live, and the audience is selected. What is missing is an eye-catching image that gets attention and gets a reader to stop scrolling.

You could commission a designer, search stock photography, or recycle an old creative. Or, inside Taboola’s campaign setup, you could type a short description and let its GenAI Ad Maker produce four images in seconds, free of charge. You decide whether to use any of them. The appeal is obvious: more options, less waiting, and lower costs. While the button is free, the media budget is not. What if consumers ignore the image? Worse, what if it looks uncanny and makes the entire ad feel less credible?

Now change seats. You are Taboola. Each candidate image costs the platform about half a cent to generate. On cost alone, rollout looks easy. But what happens after advertisers press publish? If the images attract fewer clicks than human-made ads, “free” could become expensive for advertisers and damaging to their trust. Moreover, how can the platform help advertisers better use the tool to create more effective ads? Those questions motivated “AI in Disguise—Quasi-Experimental Analysis of a Large-Scale Deployment of AI-Generated Ads,” an article that is co-authored with Yannick Exner, Jochen Hartmann, Shunyuan Zhang, and Oded Netzer, recently accepted by Marketing Science.

My co-authors and I collaborated with Taboola, whose ads appear on publisher sites such as MSN, NBC News, and USA Today. We examined the performance of 4,633 ads with more than 369 million impressions and 2.5 million clicks after the rollout of the GenAI Ad Maker in 2023.

First, the Good News: AI Ads Did Not Lose to Human Ads

We didn’t find evidence that ads with AI-generated images performed worse, on average, than their human-made counterparts. That matters because the cost difference is stark. Taboola could generate an image for roughly half a cent, while commissioned creative requires far more time and money. The result supported the continued rollout of the GenAI Ad Maker: advertisers could use the tool without accepting a detectable average click-through penalty in exchange for lower production costs.

Human-made images can look artificial, and AI-made images can look natural. The most useful question to ask is, “Does this image feel real enough to earn the click?”

What Consumers Thought They Saw Mattered More

The issue was whether the image looked as though AI made it, regardless of its origin. We recruited 612 participants and asked them to rate 1,751 images from “definitely human-made” to “definitely AI-generated.” Interestingly, they were far from perfect detectives. More than 45 percent of the AI images were perceived as human-made, while nearly one-quarter of the human-made images were perceived as AI-generated.

Those perceptions tracked performance. When both AI-generated and human-made images were seen as not looking AI-generated, the AI images performed better. As both types looked more AI-generated, that advantage shrank and could become a disadvantage. Across the ads, looking artificial was associated with fewer clicks, and the decline was steeper for images that actually came from AI.

Basically, AI performs best if it disguises itself. Here, “disguise” means avoiding visual cues that make an ad feel artificial—not hiding required disclosures or misleading consumers. That distinction matters because the look and feel of an image can be managed.

What Makes an Ad Look Like AI?

Consumers had a clear idea of what AI imagery should look like, but that idea did not always match reality. Highly aesthetic, saturated, and less warm images were more likely to be labeled as AI-generated, while realistic, clear, sharp images with larger faces were more likely to be seen as human-made.

Some of those instincts were right: the AI images were, in fact, more aesthetic, colorful, and less warm. But the biggest surprise was the mismatch. AI images were also clearer, sharper, less blurry, and more likely to feature larger faces, which are the traits consumers associated with human-made ads. In other words, some of AI’s actual visual tendencies hid in plain sight.

What Platforms and Advertisers Should Do

Design for lower perceived artificiality.

Advertisers can favor clear, sharp, realistic images and prominent faces where appropriate, while steering away from overly saturated colors and other artificial-looking cues. The platform can also incorporate these findings into both its generative creative tools and its guidance for advertisers.

Treat “looks like AI” as a quality check.

Alongside “Is the message clear?” teams should ask, “Does this image look AI-generated?” A quick pretest, or platform score on perceived artificiality, could flag risky images before launch. The source isn’t enough to determine how an audience will perceive an image. Human-made images can look artificial, and AI-made images can look natural. The most useful question to ask is, “Does this image feel real enough to earn the click?”

Further reading

AI in Disguise—Quasi-Experimental Analysis of a Large-Scale Deployment of AI-Generated Ads