900 AI ads by the numbers: what an AI production line actually looks like.
Since February 2026 our studio has scripted 2,122 ad concepts and hand-delivered 916 finished ads for DTC brands. Almost nobody publishes what happens between those two numbers — so the discourse about AI ads runs on vibes: either "AI makes infinite ads for free" or "it's all slop." Here's the actual production data. Numbers are pulled straight from our production database; the method note at the bottom explains what we counted and what we deliberately don't claim.
1. Nearly half of everything we script never ships
46%
of concepts scripted before June 15 were killed before delivery — at script review, at frame review, or superseded by a stronger variant (1,321 scripted → 715 shipped).
This is the number that separates a production model from a spam model. Generation is cheap now; the kill rate is the quality bar. Anything that reads synthetic, sounds canned, or opens with an obvious first-order hook dies before a client ever sees it. If an AI ad vendor can't tell you their kill rate, they don't have one.
2. The ramp: 5 ads to 400+ ads a month in one quarter
| Month (2026) | Concepts scripted | Ads shipped |
|---|---|---|
| February | 52 | 5 |
| March | 37 | 31 |
| April | 360 | 112 |
| May | 539 | 433 |
| June | 1,030 | 308 |
Two things worth noticing. The ramp itself: five ads in February, 433 shipped in May — the same curve any factory shows once the line stabilizes. And the June divergence: scripting nearly doubled while shipping tightened, because that's the month we hardened our review gates. Volume that outruns judgment is how feeds fill with slop; we'd rather widen the funnel's top than its bottom.
3. The anatomy of a shipped ad
8
median scenes per finished ad (5,600+ scene clips total)
117
median spoken words per ad — 80% fall between 75 and 172
12
format families in production, 50 live variants
The 117-word median matters more than it looks: that's roughly 230 words per minute of screen time — conversational pace, not voiceover-crammed direct response. AI production makes it tempting to stuff scripts (words are free); shipped ads that hold attention stay close to how people actually talk.
4. What DTC accounts actually run
| UGC | 42% |
| Tutorials & unboxings | 21% |
| Street interviews | 16% |
| Podcast & talking head | 13% |
| Statics & story-scroll | 6% |
| Animated explainers | 3% |
Share of format-tagged shipped ads. UGC leads — it's the workhorse of paid social — but 58% of shipped volume is not plain UGC. Street interviews, tutorials, podcast cuts, and animated explainers exist because UGC fatigues, and when it does, the account needs a different shape ready — not a louder version of the same one.
Method note
All figures queried from our production database on July 1, 2026. The kill rate uses only concepts scripted before June 15, so work still in flight doesn't inflate it. Word counts are extracted from shipped scripts (717 ads with parseable dialogue). Format mix covers the 616 shipped ads with format tags; earlier untagged work is excluded. What this post deliberately does not include: client ad-account performance (CPA, ROAS, spend). Those numbers belong to our clients — we walk through them live, with permission, on calls. If you quote this data, link the source and the date; it will drift as the line scales.
Want ads from this production line?
$200 per finished 30-second ad — scripts in 48 hours, delivery inside a week.
Start a brief→