Do AI-generated ads beat UGC and designer ads?

Which ad format wins on Meta now?

None wins by default. Meta's system picks the creative for each person from the pool you give it: Meta says that as businesses upload more creatives to support a diversification strategy, Andromeda works behind the scenes to pick the best one for each person (Meta, March 2025).

That makes the pool the lever: a spread of genuinely different concepts, across angles, hooks, settings and formats, gives delivery real options, while ten variations of one image give it very few, however they were made.

Where do AI-generated ads do well?

In volume and variety of product-led concepts. AI makes new settings, seasons and angles for the same SKU in days, without planning a shoot, and a testing programme runs on exactly that.

• The same product in rooms or settings that speak to different audiences
• Seasonal and promotional versions of a proven concept, without a reshoot
• Static concepts at volume, to find the angle before investing in video
• Lifestyle frames for catalogue ads (DPA), alongside the white-background shots
• Fast responses when a concept wins and needs fresh executions before it fatigues

Where does UGC still win?

Wherever the proof is a real person: a customer using the product, an unscripted reaction, a demonstration in someone's own home. That is trust AI should not manufacture, and a generated person presented as a customer is an invented testimonial.

Invented testimonials are a legal problem as well as a credibility one: the Australian Consumer Law prohibits false or misleading representations about testimonials. Use real creators for real proof, and AI for the product and the setting. More in is it legal to use AI-generated images of my product in Australia and NZ?

Where do designer-built ads win?

On offers, comparisons and brand moments, where layout and typography do the selling. A sale, a bundle, a product against the alternative or a launch announcement depends on hierarchy and copy more than on the image.

In practice the strongest statics are often both: an AI-made scene of the product, laid out by a designer with the offer, the proof point and a clear call to action. Treat AI as a source of imagery, and design as the ad.

How do we stop AI ads looking generated in the feed?

Finish them like any other ad. The AI tells that cost CTR are physical: light that does not match, a product without a contact shadow, waxy surfaces, warped text and people who look slightly wrong.

• The product exact in every frame: colour, shape, parts and label text
• Light and shadows that belong to the scene
• No generated people where a real person is the point of the ad
• A first frame built to stop the scroll, judged at phone size
• Text and offers set by a designer, never generated inside the image
• A lo-fi variant in the test, because polish is not always what wins

How should we compare AI, UGC and designer ads fairly?

On cost per winning ad, not cost per asset. A low-cost asset that never beats control costs more than an expensive one that scales, so count production and test spend together and divide by the winners each source produces.

1. Define a winner before testing, for example beating the control on CPA at a set spend
2. Test each source in the same campaign structure, with comparable budgets per concept
3. Track hit rate per source: winners divided by concepts launched
4. Add production cost and the test spend lost on losers, then divide by winners
5. Watch how long each winner lasts before fatigue, because a fast refresh changes the maths

Volume matters as much as format. On one account, an e-commerce brand Dan ran creative for, new ads went from under 5 a month to 50 to 100 a month, the ads improved as well as multiplied, and ROAS went from 2.5x to 5.5x while spend rose 300%. That is one account, not a forecast.