The short answer
AI UGC ads can spend: 77 AI-generated videos sit on our board of $50k+ Meta winners and 12 of them pass $1M in estimated spend. But only 2.6% reach the top tier, against about 10% across the board and 12.8% for real UGC. AI gets ads to scale. It rarely gets them to last.
AI ads reach the board, but rarely the top tier
Of the 610 winning videos on the Ad Radar board we tagged by production format, 77 are AI-generated: the visuals, the presenter or both were made with generative AI rather than filmed. That puts AI at about the same share as real UGC selfie ads (78). The two formats look alike at first glance. Their numbers do not.
| AI-generated | UGC selfie (real) | All tagged videos | |
|---|---|---|---|
| Ads | 77 | 78 | 610 |
| Brands | 44 | 57 | 252 |
| Median estimated spend | $229K | $592K | $318K |
| Share in the top tier | 2.6% | 12.8% | 10.7% |
| Median days live | 97 | 140 | 133 |
| Live 180+ days | 3 | 27 | 190 |
| Median length | 153s | 81s | 111s |
Two of 77 AI ads sit in Ad Radar's top tier (top 10% by winner score, live 21+ days). That is the lowest rate of any production format with a meaningful sample. Only 3 AI ads have been live for more than 180 days, against 31% of all tagged videos.
One caveat before reading too much into the days-live gap. Usable AI video is new, so no AI ad can have a two-year history yet. Part of the 97-day median is the age of the technology. The top-tier gap is harder to explain away, because the winner score also weighs variants and how fast engagement grows, not only time.
What the AI winners actually look like
The 77 ads are not the talking-avatar clips most people picture when they hear "AI UGC." Three patterns cover most of them.
Long dramatized stories
Seventeen AI ads open on a personal drama (a divorce, a cheating spouse, a run-in with an ex) and run a median of 325 seconds. The story is cinematic, voiced in the first person, and turns into a product discovery around the midpoint. One brand alone has eight of them on the board. This is where AI saves the most production cost: a five-minute filmed drama with actors would cost far more than a creator selfie.
My wife left me for the dude who trains at the gym.
- Est. spend
- $1.2M
- Days live
- 76
- Format
- Video, 509s
- Variants
- 2
One of only two AI-generated ads in the top tier. Eight and a half minutes of story before and around the product, which no creator budget would usually cover.
Narrated history and "what would happen if"
The single highest-spending AI ad on the board is a narrated history lesson about coffee. Ten other AI ads follow an "if a person did X every day for N days, here's what would happen" structure, which AI can visualize cheaply with rendered body or scene animations.
Teddy Roosevelt drank a gallon of coffee every single day.
- Est. spend
- $2.9M
- Days live
- 210
- Format
- Video, 184s
The top AI-generated ad by estimated spend, and one of the three past 180 days. A historical anecdote is easy to illustrate with AI and doesn't need a believable face.
Talking objects and voiceover listicles
Five AI ads personify the problem itself ("I'm your..." spoken by a body part or a symptom). Nineteen are voiceover listicles. Both styles use AI where it is strongest: illustrating something a camera can't film.
Never ignore these three hidden distress dog behaviors: paw licking, itching, and head shaking.
- Est. spend
- $520K
- Days live
- 110
- Format
- Video, 50s
The other top-tier AI ad. Fifty seconds, three named behaviors, and it sends traffic to a quiz. AI handles the visuals, the structure does the selling.
Where AI hurts: the fake person
The pattern is consistent. AI ads that work illustrate a story, a mechanism or a scene. AI ads that imitate a real customer talking to a phone have a harder time.
Real UGC wins on believability: a person with a real bathroom, real skin and a real dog. Our UGC ads examples show 47 of 78 real UGC winners using the same simple bridge from personal problem to solution. That bridge depends on the viewer believing the person had the problem. An AI face weakens exactly that link, and the top-tier gap (12.8% against 2.6%) is what that looks like in the data.
There is also a policy point. An AI-generated person presented as a customer giving a testimonial raises the same questions as a fabricated review. Check Meta's ad policies and the FTC endorsement rules for your category. That responsibility sits with the advertiser.
Who uses AI-generated ads
AI ads skew toward cold audiences and supplements:
- Awareness: 48% target unaware viewers, against 39% of all tagged videos. AI is mostly used to tell stories to people who don't know they have a problem yet.
- Product form: 39 of 77 are supplements, 12 are devices.
- Niches: skin (16), digestion (15), pets (13), stress and mood (10).
- Landing: 38 go to a product page, 22 to a dedicated landing page.
The concentration is high. 44 brands made the 77 ads, and two brands account for 16. When a few brands make most of a format's ads, their testing habits shape the numbers, so read the AI figures as a pattern across 44 brands, not a law.
Where AI helps in the ad process
The data argues against replacing your best human formats with AI clones. It does not argue against AI. Here is where it pays off for the brands we track and for creative teams in general.
| Job | Use AI? | Why |
|---|---|---|
| Visualizing a mechanism or history | Yes | Cheap scenes a camera can't capture. The top AI ads by spend are here. |
| Long story ads with actors | Test it | Cuts the cost of a 5-minute drama. Expect shorter runs. |
| Fake customer testimonials | No | Lowest believability, policy exposure, weakest data. |
| Scripts and hooks | Yes | Draft 20 openers from proven patterns, then film the best with a person. |
| Variants of a proven ad | Yes | New first lines, captions, cuts on a body that already wins. |
| B-roll and inserts | Yes | Fill gaps in a filmed ad without a reshoot. |
The best use of AI on the numbers is upstream of the camera: research, scripts and variants, then a human on screen. We cover the script side in AI ad script writing and the tooling in AI ad creative tools.
A practical way to test AI UGC
- Start with a format AI is good at. Story narration, a "what happens if" visual, a personified problem. Not a fake reviewer.
- Borrow a proven structure. Pull the hook pattern and angle from human ads already winning in your niche. On the Ad Radar board you can filter by niche and production format and read the transcripts.
- Make the ad long enough. The AI winners run a median 153 seconds. A 15-second AI clip has little to hold attention with.
- Judge it at 30 days, not 7. The weakness of AI ads is staying power. An AI ad that holds up past a month is the exception worth scaling.
- Film the winner with a person. When an AI concept wins, reshoot it with a real creator or founder and run both. The human version has the better odds of reaching the top tier.
If you work with Claude, the Ad Radar MCP server can pull the transcripts of winning ads by format so the model drafts from what is spending, not from generic patterns. See Claude prompts for Facebook ads for prompts that work with it.
Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.