Every media buyer can recite “creative is the new targeting”. Far fewer have changed how the ads actually get made. The monthly batch still looks like 2023: one proven concept, eight near-identical cuts, a new hook bolted on the front, maybe a different colour grade. If that’s your account, the phrase hasn’t reached the part of the operation it was aimed at.
Meta announced Andromeda in December 2024. It’s the retrieval engine at the front of the auction, sifting tens of millions of candidate ads in milliseconds with roughly 10,000 times the model capacity of what it replaced. It decides which ads are even plausible for a person before the auction runs, and it makes that call by reading the ad. Who’s on camera, what problem the first line names, the format, the claim, the tone. Those used to be creative decisions taken after targeting was set. They are the targeting now. Meanwhile interest stacks and lookalike ladders became suggestions under Advantage+, and you can’t out-segment a model that builds its own segments from live behaviour.
“Make more ads” is the expensive misreading
It feels like action, which is what makes it costly. If the whole batch speaks to the same person about the same problem in the same shape, you’ve rebuilt one narrow interest stack with more files in it. Meta folds near-duplicates into a single delivery entity. Volume without distinctness is one ad, uploaded 30 times.
Targeting lives in three places inside a UGC ad. Persona is the strongest: a 26-year-old filming a get-ready-with-me reaches a different corner of the auction than a 55-year-old talking about knee pain on a morning walk. Same product, different ad populations. Angle is second, because the problem your opening line names tells the model what intent to match. “I stopped hitting snooze”, “my sleep tracker finally went green” and “cheaper than my old routine” are three audiences wearing the same product. Format is third and the most underrated. Talking head, voiceover B-roll, side-by-side demo, static with a review quote. Format decides who stops scrolling, and the system learns from who stops.
The test before a batch ships: strip the logo and see whether a stranger can sort the ads into separate “this one is for X” piles. One pile means you made variants, not diversity.
Build the brief as a grid
Rows are personas, pulled from reviews, support tickets and post-purchase surveys in the customers’ own words. Not marketing personas with stock-photo names.
Each persona gets two or three angles, written as the actual first line the viewer hears. “Energy” is a theme. “I survived the 3pm crash without a third coffee” is an angle.
Assign formats on purpose so neighbouring cells never share a shape. Not every cell needs video.
Cap variations of a winner at two, one new hook and one new cut. Fifteen crops collapse back into one entity anyway.
Hold a quarter of the slots open and fill them mid-month with iterations on whatever the live grid is rewarding.
A 12-slot month for a sleep supplement brand, nine committed and three held back:
| Persona | Opening line | Format |
|---|---|---|
| New parent, 28-38 | “I survived the 3pm crash without a third coffee” | Creator talking head |
| New parent, 28-38 | “The wind-down routine that finally stuck” | Voiceover B-roll, evening routine |
| New parent, 28-38 | “Cheaper than the sleep stack I was buying” | Static, price breakdown and review quote |
| Shift worker, 25-45 | “Day sleep is a different sport. This is what worked” | Talking head, blackout-blind room |
| Shift worker, 25-45 | “What I take at 7am after a night shift” | POV demo, morning-after framing |
| Shift worker, 25-45 | “My partner sleeps 9 to 5. I don’t” | Static, split image with caption hook |
| Desk worker, 40+ | “I wake at 3am and my brain starts a meeting” | Older creator, talking head |
| Desk worker, 40+ | “My watch’s sleep score, 30 days before and after” | Screen recording and voiceover |
| Desk worker, 40+ | “I tried magnesium, teas and podcasts first” | Voiceover B-roll, failed alternatives |
| Open x3 | Iterate whichever cells the account rewards | Match the winning cell’s format |
Then read the results by row, not by ad. Which persona-and-angle cells found cheap buyers and which found nothing. Next month doubles down on the live cells and swaps the dead ones. That habit compounds faster than any single winner.
How much volume
One post-Andromeda dataset put brands testing 20+ new ads a month at 65% higher ROAS than brands testing under 10. Directional numbers from one source, but they match practice: accounts that keep feeding the system get cheaper, accounts that ration creative get repriced.
Twenty ads only count as twenty if they’re distinct, and the right number scales with spend. A $5k account doesn’t need forty. Fatigue is the other half of the question. Refresh windows that ran six weeks now close in two to three, which is what the open slots are for.
Stop doing these
Rebuilding interest stacks when performance dips. The model overrides them, and a dip is a creative signal problem now.
Running the 1%, 3%, 5% lookalike staircase. It splits your conversion signal across pots and gives the algorithm less to learn from in each.
Duplicating ad sets to reset delivery. You’re resetting learning, not luck.
Cutting one winner into a dozen crops and calling it testing. That’s a rendering job.
The hours land somewhere better. Review mining becomes a monthly ritual instead of a launch-week one, because the grid eats fresh angles. The creator brief becomes something you manage, a roster of personas your ads actually cover rather than one house style repeated. And the weekly review moves from “which ad won” to “which cell won”, which is the question that tells you what to make next.
The uncomfortable part sits underneath all of this. A persona and angle grid is easy to write and hard to feed. Twelve to twenty distinct creatives a month, every month, with the fatigue window closing at three weeks, is a production problem long before it’s a strategy problem.
That’s the part I run. New concepts land every Friday, each winner gets capped at a couple of variations instead of fifteen crops, and the grid gets re-read monthly so dead cells get swapped rather than refilled. If your brief has been a list of concepts rather than a grid, book a free call and we’ll build one against your account.
Volume and fatigue figures compiled from published 2026 post-Andromeda datasets. Andromeda capacity figures from Meta’s own engineering announcement, December 2024.