Attention Without Conversion: The First Field Experiment on AI Video Ads Is In

Attention without conversion: the number is finally in

A peer-reviewed field experiment has put numbers on the promise and the limit of AI video ads: they win clicks, hold viewers longer and draw fewer negative reactions than human-made ads, and they still convert fewer customers. The attention without conversion problem is now measured evidence, not anecdote, and it redraws where commercial video teams should deploy generative creative.

The study, by Huimin Li and Qing Li of Hong Kong Baptist University, was presented at the 48th Annual ISMS Marketing Science Conference in June 2026. It compares AI-generated and human-created video creatives across the advertising funnel using more than 150 unique ads in a randomized field experiment. On the upper-funnel outcomes teams usually celebrate, the AI-made ads won across the board: higher click-through rates, more effective viewer retention and fewer negative reactions from the audience.

The picture inverts downstream. Relative to the human-created control ads, AI-generated creatives were less likely to generate conversions, even while outperforming on every attention metric measured above the click. The authors call the pattern a tension between attention efficiency and persuasive effectiveness, and they locate the failure precisely: attention alone is insufficient to deliver conversion when trust, credibility or purchase intent matters.

For commercial video teams this is the most consequential finding since creative parity studies claimed generative direct-response work matched human output on both CTR and conversion. Both results can be true. Parity describes what well-executed AI creative can achieve in aggregate benchmark data; the experiment isolates what happens inside one controlled comparison, where the attention metrics and the revenue metric moved in opposite directions. When a dashboard splits that way, the routing decision becomes a production decision.

The upper funnel is where generative video wins

The upper-funnel wins deserve to be taken seriously, because they are exactly the outcomes generative tools were built to optimize. A model trained on millions of high-performing clips is very good at hooks, pacing and pattern interrupts, and the experiment shows that competence carries into live delivery: AI-made ads earned higher click-through rates and held viewers more effectively than the human-created comparison set, with fewer negative reactions signaling less audience irritation along the way.

The behavioral backdrop was already pointing this way: {{link}} documented views climbing while watch time and engagement per video fall across social platforms, and the experiment adds a creative dimension to that drift. Attention is getting cheaper to buy and harder to keep, and generated creative is very good at winning the first exchange.

The trap is reading those green arrows as proof the creative works. Every metric the AI ads won sits above the click; every margin the human ads held sits below it. A media plan that optimizes only what the experiment calls attention efficiency will keep scaling the half of the funnel where AI wins and never notice the half where it loses.

The behavioral backdrop was already pointing this way: the video attention split documented views climbing while watch time and engagement per video fall across social platforms, and the experiment adds a creative dimension to that drift.

Why attention stalls before purchase

Why does attention fail to convert? The authors point at trust, credibility and purchase intent, the properties a viewer evaluates after the hook lands and before the checkout opens. A generated clip can be gripping and still leave the viewer unsure the product is real, the claim is honest or the brand is accountable for what the ad shows.

That is consistent with what {{link}} has already mapped: skepticism toward AI-made creative concentrates in exactly the younger cohorts advertisers most want to reach, and it attaches to the brand rather than to a single campaign. A viewer who enjoys the ad and distrusts the advertiser produces precisely the funnel shape the experiment measured.

There is also a sameness problem compounding the trust problem. Practitioners quoted in industry coverage of the research report that fully AI-generated videos now face lower CTR precisely because audiences see so much similar content, while videos that combine real human characters with AI elements perform better. Audience fatigue with the same face, the same lighting and the same impossible camera move erodes the credibility that conversion requires, even when the viewer cannot name what feels wrong.

That is consistent with what the consumer trust gap has already mapped: skepticism toward AI-made creative concentrates in exactly the younger cohorts advertisers most want to reach, and it attaches to the brand rather than to a single campaign.

What 16 billion display impressions add

A September 2026 working paper from Columbia Business School and partner institutions adds the perceptual mechanism. The researchers analyzed more than 16 billion display ad impressions and 116 million clicks from over seven thousand advertisers across nearly 50 product categories, and found that ads with AI-generated images beat human-generated images on click-through, but only when the images did not look like AI.

The visual features cut both ways. AI tools generate more aesthetic images with larger, clearer faces, and consumers read those traits as signs of human craft. Intense color saturation, by contrast, signals AI generation to consumers and suppresses clicks. For video teams the implication is uncomfortable but useful: the qualities that make generated output look impressive in a review queue, hyper-saturated grading and over-polished aesthetics, are the same qualities audiences read as artificial. Restraint is not a taste preference here; it is measurable engagement.

Two caveats keep the finding honest: the dataset is display advertising rather than video, and clicks are not sales. But the mechanism travels. If perceived artificiality taxes engagement even at the impression level, it plausibly taxes persuasion harder downstream, where money changes hands. The same production discipline, natural light, real faces, restrained grading and nothing uncanny, protects both stages at once.

Route AI creative by funnel stage

The routing rule that falls out of the evidence is simple: deploy AI video where attention is the product, and keep humans in the frame where persuasion is. Top-of-funnel reach, variant volume for testing, localized cutdowns and iteration speed are the jobs generated creative demonstrably does well. Used there it compounds cheaply, which is the whole logic of {{link}}, and the experiment now supplies the boundary: the stage where it stops paying.

Conversion-stage creative is the other case. Product demos, objection handling, testimonial and proof formats all lean on credibility the audience does not grant a synthetic presenter by default. Practitioner evidence in the same coverage points at the hybrid pattern: real human characters carrying the persuasive load, with AI handling scale, variations and production polish around them.

The cheapest place to start is pre-ship measurement: {{link}} gives teams a repeatable way to check whether a cut actually moves feeling before spend finds out. Pair it with a funnel-stage tag on every asset so the dashboard can separate attention wins from persuasion wins instead of blending them into one misleading average.

Used there it compounds cheaply, which is the whole logic of the awareness economics, and the experiment now supplies the boundary: the stage where it stops paying.

The cheapest place to start is pre-ship measurement: emotion testing gives teams a repeatable way to check whether a cut actually moves feeling before spend finds out.

Measure persuasion, not just attention

Measurement is the other half of the fix. CTR measures clicks and ROAS measures attributed spend, and neither alone tells the story the experiment tells. Teams running AI video should read the funnel in stages, hook rate and retention for the top, conversion and incremental revenue for the bottom, and compare AI-led, hybrid and human-led creative within each stage rather than across all of them.

The experiment also reframes what counts as a win. Comparable performance at substantially lower cost is a legitimate commercial result, and it does not need to beat human work on every metric to justify deployment. What it does need is honest staging: know which outcomes a generated asset was built to deliver, judge it on those, and record the boundary where a human-led cut took over.

The green arrows on the dashboard are real. The experiment simply shows they point at attention, not at revenue. Teams that route by funnel stage, keep humans where trust is priced in, and measure persuasion where the money changes hands will get both, and the teams that only optimize the click will keep paying for the difference.