the ai ugc playbook

may 2026

ai ugc is the most overhyped, under-systematized growth channel of the last 18 months. everyone is shipping the same talking-head avatar saying the same script over a stock b-roll loop. it converts for about four weeks, then the algorithm sniffs the pattern and the cpms collapse.

i ran an ai ugc operation that hit 40k mrr in under two months. then i watched it die when i stopped treating it as a content engine and started treating it as a creative one. this is what i learned.

what ai ugc actually is

ai ugc is short-form video that looks like a real person talking to camera but is partly or entirely synthetic — generated voice, generated face, ai-edited cuts, ai-written hooks, often ai-selected b-roll. the user-generated part is the aesthetic, not the source.

it sits between three older categories:

paid ugc — you hire a creator to film an ad. expensive, slow, hard to iterate.
stock-ad creative — generic b-roll plus voiceover. cheap, looks generic, low ctr.
organic creator content — a real person posting from their real account. high ceiling, no control.

ai ugc is the fourth thing: the cost structure of stock-ad, the aesthetic of paid ugc, the iteration speed of code. that's the entire pitch. that's also why most of it is bad — people optimize for the cost structure and forget the other two.

the stack i used

no part of the stack matters as much as people pretend. the moat is in the workflow, not the tools. but for the record, the v1 was:

hook bank — a google sheet of 400+ hooks i'd manually pulled from top-performing organic creator content in adjacent niches. tagged by emotional register (curiosity, fear, status, contrarian, vulnerability).
script gen — a custom gpt that took a hook + a product spec + 3 example scripts in the brand voice and produced 5 variants. nothing fancy. the prompt was 2k tokens.
avatar — heygen for v1, then a fine-tuned hedra setup for v2. heygen was faster to start, hedra produced output that didn't look like heygen, which was the entire point.
b-roll — a tagged library of ~3,000 clips i'd scraped and license-cleared, indexed by clip subject and energy level.
edit — capcut with a templated timeline. no fancy ai editor. an editor in the philippines doing 30 cuts a day was faster and cheaper than any tool i tested.
distribution — 12 tiktok accounts and 8 instagram accounts running the same creative with different first-frame variants.

total monthly cost at peak: about $3,400. revenue at peak: 40k mrr.

three case studies

1. a sleep supplement, $0 to $18k mrr in 5 weeks

the brand had a generic landing page and a saturated category. we ignored the supplement entirely in the first 2 seconds of every video and led with the symptom — a girl in bed at 3am scrolling, a guy waking up tired before his alarm. the avatar said one line: "i didn't realize my sleep was bad until i tracked it for two weeks."

every variant was the same script with a different first frame and a different opening adjective. we ran 80 variants in week one. four hit. those four ran for the rest of the cycle.

learning: the creative is the targeting. you don't pick an audience, you pick a hook, and the algorithm finds the audience for the hook.

2. an ai writing tool, killed at $3k mrr

we built a beautiful pipeline. avatar that looked exactly like a midwestern engineer. scripts that name-checked vs code and notion. it converted for two weeks. then it didn't.

the problem: the audience for ai writing tools is too sophisticated. they recognized the avatar as ai by week three. comments turned hostile. the algorithm down-ranks creative with negative sentiment in the first hour.

learning: ai ugc dies fastest in audiences that are themselves familiar with ai. it works best in categories where the buyer is buying a result, not a technology.

3. a beauty brand, $0 to $14k mrr, then plateau

we did everything right. 60 hook variants, three avatars, brand-matched palette. it scaled clean to 14k. then we couldn't push past it.

the bottleneck wasn't creative — it was creative diversity. all our winners had the same emotional register (gentle, reassuring). we'd taught the algorithm we were one kind of brand. the next pool of viewers wasn't responsive to that register.

learning: scale demands creative range, not creative volume. ten emotionally distinct hooks beats a hundred variations of one.

why most ai ugc fails

three failure modes, in order of how often i see them:

the creative is generated, the strategy isn't. people use ai to make the videos and a checklist to decide what to make. it should be the inverse — humans pick the angles, ai produces the volume.
no first-frame discipline. the first frame of a tiktok is the entire product. ai ugc operators iterate the script and re-use the same opening shot. that's why their cpm collapses.
the avatar is the message. if the audience leaves your video remembering the avatar, you've lost. the avatar should be invisible. the message should be the message.

what's next

the half-life of any ai ugc workflow is about six months. by month seven the platforms have updated their detection, the audience has updated their pattern matching, and your competitors have copied your stack.

the durable thing isn't the tooling. it's the rate at which you can replace the tooling. operators who can rebuild their pipeline in two weeks survive. operators with one beautiful pipeline don't.

i'm building the next version of this inside warmr — distribution infrastructure across 50+ physical iphones, because the platforms are starting to penalize content posted from data centers. the creative stack is just the input. the distribution stack is what compounds.

if you're running ai ugc and want to compare notes, email me. i answer everything.