Direct answer: The best-documented gains from AI in social advertising come from automated delivery (targeting, placement, bidding), where Meta reports small but consistent conversion lifts. AI-generated creative is the weaker, riskier half: one field study found a click-through gain for fully AI-made ads, but the same study found that labeling them as AI cut clicks by almost a third.
Meta’s third-quarter 2026 report gives a clear example of how the topic gets muddled. Ad revenue rose 26% to $63.1 billion, and the coverage credited “AI.” Check the components: impressions grew 10%, and the average price per ad grew 14%. Multiply them (1.10 × 1.14 = 1.254), and you get roughly 25%, which matches the reported 26% within rounding. Revenue growth came from showing more ads at higher prices. Meta did not disclose a dollar figure for AI-attributable revenue.
That is the first thing to understand before you spend a euro on “AI ads.” The phrase covers two different products with different evidence.
Two products sold under one name
AI-delivered ads use machine learning to decide who sees an ad, where, and at what bid. You cannot see this layer, and neither can your customers. AI-made ads use generative models to produce the image, video or copy. Customers can see this layer, and increasingly, regulators and platforms want them to be told about it.
The figures in the diagram come from the sources discussed below. The ranges are not like-for-like. One side is conversions reported by the platform, the other is clicks measured by academics, and they should not be added together.
The delivery layer: modest, real, and hard to audit
Meta’s Advantage+ suite bundles audience, placement and budget automation. Meta said it passed a $75 billion annual revenue run rate in the second quarter of 2026. The third-quarter 2025 figure was $60 billion. A run rate is an annualized snapshot, not booked revenue, and it measures how much advertisers spend through these tools, not how well the tools perform.
Performance is where the numbers get small. On the Q3 2026 call, Meta reported that its GEM ranking model lifted conversions 1.7% for off-site conversion campaigns and 9.6% for app ads, and that its new Generative Recommender drove a nearly 2% increase in conversions across targeted segments. These are first-party claims from the seller. They are plausible, and they are also the best evidence available, but no independent audit sits behind them.
A technical distinction matters here. GEM does not generate ad creative; the “generative” refers to predicting which ads will perform for which users. (This description comes from an ad-tech vendor’s blog, a secondary source, so treat the retrieval details, such as the shortlist of roughly 500 candidate ads, as approximate.) Most “AI ads” headlines blur this, which leads advertisers to expect creative miracles from a system that mostly does matching.
What the 14% price rise means for you
Meta’s reported 14% rise in average price per ad is a platform-wide average, not your CPM. But it shows the direction. A simple illustration with assumed inputs: if your cost per thousand impressions rises 14% and AI ranking improves your conversion rate per impression by 2%, your cost per acquisition changes by 1.14 ÷ 1.02 − 1 ≈ +11.8%. To hold CPA flat, you would need a conversion lift near 14%. The small lifts Meta reports do not automatically pay for higher auction prices. Whether they do for your account is a question only your own holdout test can answer.
My position: the delivery layer is worth using, and for small advertisers it is hard to avoid, since Meta is making Advantage+ the default for new campaigns. The advertiser’s real leverage is feeding it clean conversion data. Pixis argues that teams with complete conversion signals get better automated results than teams spending the same with gaps. That is a vendor’s inference, but the mechanism follows directly from how optimization systems learn.
The creative layer: the best evidence, and the biggest trap
The strongest independent study here is “The Impact of Visual Generative AI on Advertising Effectiveness” by Anindya Ghose, Hyesoo Lee, Vilma Todri and Panagiotis Adamopoulos (NYU Stern and Emory, SSRN, November 2025). It compared three kinds of visual ads: human-made, human-made then modified by generative AI, and fully AI-generated. Three findings, in order of importance for practitioners:
- Fully AI-generated ads raised click-through rates by up to 19% versus human expert ads. The field test ran on Google’s Display Network.
- Using AI to modify existing human ads showed no significant improvement. Forcing the model to work inside an existing design limited it.
- Telling consumers an ad was AI-made cut click-through rates by 31.5%. In absolute terms that is about 1.17 percentage points, according to a secondary summary of the paper.
The non-obvious part: the hybrid workflow is the worst of both
Most teams use AI the third way, as a polishing step on human work: swap a background, adjust lighting, retouch. The study’s evidence says that is where you pay the cost (tool time, labeling exposure) with no measured gain. The result favors an either/or choice per format: generate from scratch where the stakes are low and volume matters, or keep human production where brand precision matters. A retouch pass sits in the middle and has no support in this data.
This applies to static visual ads. It says nothing about copy, video, or audio, and the field test was display advertising, not a social feed. The authors measured clicks, not purchases. A higher CTR can still produce lower sales, so do not read “+19%” as “+19% revenue.”
The disclosure bill
The 31.5% figure is real but narrow. Consider what else is known:
- IAB’s AI Transparency and Disclosure Framework (first published January 2026, updated to version 2 in August) recommends disclosure only when AI materially affects authenticity, identity, or representation, such as synthetic humans, realistic generated scenes, and chatbots posing as people. Using AI to draft copy sits in a different category from fabricating a realistic video of a person.
- Its companion survey found that 82% of ad executives believe Gen Z and Millennials feel positive about AI-generated ads, versus 45% of those consumers who actually do. The gap widened from 32 points in 2024 to 37 in 2026. This is an industry trade body’s research, commissioned from Sonata Insights.
- Meta now requires disclosure in certain cases of AI-generated content in ads, according to an agency summary. Check Meta’s current ad policy directly, since that summary is secondary and policies change.
- The EU AI Act’s Article 50 targets deepfake-style content that imitates real people, objects or events in a misleading way. It does not automatically require labels on every AI-produced background or product render, per one legal-adjacent commentary. If you advertise in the EU, confirm current application dates and scope with counsel.
The practical reading: you will probably be required to label the riskiest creative (realistic people, realistic events), and the penalty will land where trust matters most. That favors using generated imagery for products, abstract scenes and illustrations, and avoiding fabricated humans.
What to do on Monday
- Split your account into two tests. Run delivery automation (Advantage+ vs. manual structure) as one experiment, and creative origin (human vs. fully generated) as a separate one. Never change both at once; you will not be able to attribute the result.
- Use a holdout, and judge on cost per incremental conversion, not CTR. Meta’s own figures are conversion lifts under 10% for the components it quantified. If your test cannot detect a difference of that size, you do not yet have evidence either way. Run it long enough to reach several hundred conversions per arm; if your volume cannot support that, say so rather than guessing.
- Stop the hybrid middle. Pick formats where you will generate from scratch and formats where you will not. Do not pay for AI retouching of human work on the strength of the current evidence.
- Test a labeled arm where disclosure applies. If you must label, find out your own penalty. The 31.5% is one study’s average; yours could be larger or smaller.
- Fix your conversion signal before you touch creative. Server-side events, deduplication, and correct value passing give the delivery system something to learn from.
If the delivery test shows no lift after the holdout period, revert to your manual structure and spend the savings on creative volume instead. The creative test has no downside reference yet in this data, so treat any win as provisional until you repeat it.
The question the coverage skips: will Meta really automate everything?
The WSJ reported in June 2025 that Meta aimed to let brands supply a product image and budget, with AI producing the ad and handling targeting by the end of 2026. As of this writing, I could not confirm a general launch of that URL-plus-budget product. What exists is the creative toolkit that nearly 10 million advertisers used at least one tool from in Q3, plus automated delivery. Check Meta’s Ads Manager release notes before you plan around a full-automation date.
Whether it launches or not, the implication holds. If every advertiser runs on the same delivery system, targeting stops being a differentiator, and creative and first-party data become the only variables you control. Large brands remain reluctant: agencies report clients still reject fully automated concepts. That tells you the market is not yet at the point where automation substitutes for taste.
Evidence ranking
| Claim | Source type | Strength |
|---|---|---|
| Fully AI-generated visuals raised CTR to 19% (display) | Independent field study, working paper | Strong, but one study, CTR only |
| Disclosure cut CTR by 31.5% | Same study | Strong, same limits |
| Advantage+ run rate above $75B | Meta earnings call, press-reported | Reliable as a spend figure, not a performance figure |
| GEM +1.7%/+9.6% conversions | Meta, press-reported | First-party, unaudited |
| 82% vs 45% perception gap | IAB-commissioned survey | Useful, from an interested trade body |
| Fully automated ads by end-2026 | WSJ report via secondary coverage | Unconfirmed as launched |