What the best measurement shows
A peer-reviewed marketing-science study published in August 2026 compared AI-generated ads with ads made by a human designer in a live Instagram campaign, at a company selling leisure activities.
The result went to the AI: a 0.98% click-through rate for the AI creative, 0.65% for the human designer. In 99.59% of the resamples, the AI batch beat the designer's.
The important part, though, is the follow-up. Eighteen months later, with no retraining of the model, in peak season: 3.38% for the AI portfolio, 3.24% for the human campaign. Both got much better — and the gap all but disappeared.
Two things follow from that. AI creative really does work, it is not marketing talk. But its advantage is not a permanent property: it goes stale if you do not retrain, and good human work catches up.
Where the real gain is
Not in the prettier picture. In the fact that twenty creatives take the same time as two.
That matters because most of ad performance does not hinge on execution, it hinges on how many different claims you tried. Twenty creatives are not twenty prettier pictures: they are five claims, each in four visual treatments — and those five claims are what you are actually testing.
THE STRUCTURE OF THE SERIES
Layer 1 - HOOK (the first 5 words of the copy): 5 claims
price / time saved / risk / status / "not for you if..."
Layer 2 - VISUAL: 4 types
product alone / in use / before-after / text only
Layer 3 - FORMAT: 2 ratios (portrait / square)
5 x 4 = 20 creatives, in 2 ratios.
One hypothesis = one line in the hook. If you cannot say
WHAT a variant claims, it is not a variant, just another picture.
The visual layer runs on the same workflow that producing social visuals at scale does: a template, fixed brand elements, an interchangeable layer. In organic content the same layered thinking shows up in the calendar — the monthly content calendar is the post that covers that side.
What never to leave to the machine
Three things in the series cannot be automated.
- The hypothesis list. The machine varies, it does not think. If all twenty creatives say the same one thing in different words, you are running one test instead of twenty — only more expensively.
- The factual claim. The price, the deadline, the guarantee, the words "market leader": nobody checks those for you. In generated copy these are the sentences that slip through most often, and they are the most expensive ones.
- The recognisable person. Using a real person's face or voice without their permission is outright prohibited on most platforms, not merely subject to labelling.
Labelling is no longer optional
This is what changed in 2026, and anyone carrying on with the old practice runs into ad rejections.
Google. From July 2026 a two-state setting in Google Ads marks whether an asset was made or edited with AI. Creative made with Google's own generative tools is labelled automatically, and the advertiser cannot override that label. For campaigns targeting the European Union, India and New York, labelled assets also carry a visible notice. A separate panel shows the user how the ad was made. Google itself spells out that using the setting does not guarantee legal compliance.
TikTok. Significantly edited and AI-generated content is allowed — but only with a label. If the platform identifies unlabelled AI content, it rejects or restricts the ad. Fully generated image, video and audio need a label, as does a person doing or saying something they did not actually do. Lighting, colour and background removal do not.
Meta. Ads made with its own generative features are labelled automatically, and it also tries to detect the ones made with third-party tools. On photorealistic images of people the label goes somewhere visible.
YouTube. Disclosure is mandatory for realistic AI content, and from May 2026 automatic detection runs as well. One thing is worth knowing, because it is why many people put this off: YouTube states that the label by itself does not affect how the system recommends the video, nor whether it can be monetised.
On the legal background behind all this and on the technical side of marking, labelling AI video is the post that goes into the detail.
Political and social-issue ads: a separate world
If the ad is about an election, a candidate or a social issue, the rules above are not the ones that apply — stricter ones are. The EU political advertising regulation has applied in full since October 2025: a political ad has to be clearly marked, it has to state who paid for it and how much, and where targeting is used, the targeted audience as well. Separate, explicit consent is required for political targeting, and profiling built on special categories of personal data is prohibited.
The platforms maintain separate mechanisms for this as well. If your activity falls into this bracket, do not settle it with the general AI label.
Creative fatigue: the number that gets misquoted
The most frequently quoted figure comes from one platform's own analysis: click-through rate degrades with the number of prior impressions, and by the fifth impression it is roughly half what it was at the first. The same analysis measures that more than 19% of ad impressions had already been seen more than five times by the same user, and that swapping out fatigued creative improved conversion rate by 8% on average.
Except there is a much larger independent analysis too: 2.8 billion impressions, thirty campaigns. There, ten of the thirty campaigns show no detectable fatigue at all, and only four show it strongly.
So the right reading is not "swap it after the fifth impression". It is: fatigue varies enormously from campaign to campaign, so measure your own. Looking at frequency and click-through rate together tells you whether you have any. If you do not, creative thrown out every week is just money thrown out.
How to decide the winner
Here comes the hardest constraint: you cannot test twenty creatives cleanly against each other. Every further variant raises the chance of finding a winner by accident, and the price of correcting for that is steep.
In practice the process therefore runs in two rounds: the twenty creatives are exploration, from which the platform's algorithm and the raw numbers weed out the obviously weak ones, and only the best two or three go into a clean, pre-registered comparison. The sample-size mathematics is the same as for copy variants — the details are in testing landing page variants, a post of its own.
Common misreadings
- "AI creative always performs better." In one peer-reviewed measurement, yes — but the advantage melted from 0.33 percentage points to 0.14 over eighteen months.
- "Labelling hurts your results." YouTube says explicitly that the label by itself does not affect recommendation or monetisation. A rejected ad, on the other hand, definitely does not perform.
- "It is enough not to mention it." The large platforms run automatic detection, and rejection is not a theoretical risk.
- "Twenty creatives are twenty times the information." Twenty creatives are twenty guesses if they all claim the same thing. The value sits in the different claims.
- "You have to swap after the fifth impression." That is one platform's own curve. On a larger sample, a third of campaigns have no fatigue at all.
What to take away
Producing at volume pays off not because it gives you prettier creative, but because it lets you try more hypotheses in the same amount of time. That takes a hypothesis list, a labelling routine, and a decision rule that applies to two or three variants rather than twenty. The machine took over the varying — not the thinking and not the responsibility.