Meta Andromeda for creative teams: what Meta said, what to brief
What Meta has actually said about Andromeda, which popular claims hold up, and how to brief and judge concepts that are genuinely different.
Andromeda is Meta's ad retrieval system. It's the first stage of delivery: for each person it narrows tens of millions of candidate ads to a few thousand, and ranking models then pick what they actually see. Meta hasn't set a weekly number of new ads, or said how similar is too similar. What it has said, in December 2025, is that ads that look or feel alike are treated as variations of the same creative, and what it learns about delivery is shared across them.
So the brief is easy to state: more concepts that are genuinely different (a new persona or a new reason to buy) and fewer versions of last month's winner. Where teams go wrong is in reading their own account. They count ads instead of concepts, compare results that don't compare, and follow rules Meta never published.
What Meta has said about Andromeda
Meta's engineers described Andromeda in December 2024 as retrieval, "the first step in our multi-stage ads recommendation system": it picks from "tens of millions of ad candidates" and keeps "a few thousand". The post gives advertisers no advice on hooks, similarity or how many ads to run.
The split between retrieval and ranking is what most explainers miss. Retrieval decides whether your ad makes someone's shortlist. Ranking models, trained by Meta's GEM foundation model, decide whether it wins. So a distinct ad should have a better chance of making the shortlist, but diversity won't rescue a weak ad.
Meta's 2024 post said retrieval worked from precomputed ad embeddings: numbers describing each ad, worked out ahead of time. Now language models "read each ad's text and images" (October 28, 2026), and Meta combines that with what it has learned from how people engage, into tokens where "Similar items share tokens" (October 19, 2026). That's how a new ad can benefit from what the system already learned about similar ads.
Selection changed too. Retrieval now predicts the ideal ad for a person and "matches that to the ads available in our system" (October 28, 2026).
| Date | Source | What Meta said |
|---|---|---|
| Nov 19, 2024 | Meta for Business | Keep "at least 20 diversified ads" in Advantage+ shopping campaigns; its best performers launch new ads weekly. No mention of Andromeda. |
| Dec 2, 2024 | Meta Engineering | Andromeda is the retrieval stage, built for far more creative. |
| Apr 22, 2025 | Meta for Business | The focus "has shifted from niche targeting to creative diversification". Asks for different themes, messages and visuals. No number of ads. |
| Dec 16, 2025 | Meta for Business | The ads system "groups together ads that share key visual and thematic attributes" and treats them as "variations of the same creative", so "learnings and delivery optimizations are shared". |
| Jul 29, 2026 | Q2 2026 earnings call | A Generative Recommender "using LLMs to reason about ad content and user preferences together", and the first generative model in ads retrieval. |
| Aug 2026 | Ads Manager, reported by Search Engine Land on Aug 26 | A Creative diversity column rates an ad set's visual diversity as Low, Medium or High. Labeled in development, with no published thresholds. |
| Oct 28, 2026 | Meta Engineering and Meta for Business | A rebuilt Andromeda "scores every eligible ad"; GEM "teaches the models" that rank ads; language models "read each ad's text and images". No number of ads or similarity limit. |
The December 2025 post is the most useful thing Meta has published for creative teams. It says "Creative diversification and creative iteration are two separate concepts." Its example of iteration is "two ads with identical visuals, but different text CTAs". Its example of diversification is a Reels ad for Gen Z next to a creator partnership ad for new moms, with a different audience, format and face on screen.
Myth check: popular Andromeda claims against Meta's sources
Only one of the most repeated Andromeda claims has Meta's backing: that its ads system groups look-alike ads. Meta has published no similarity cap or weekly quota, and nothing saying Andromeda reads hook rate. Several pages below are from tools that compete with Yolodex. "Not supported" means Meta hasn't published anything that backs a claim, not that it's been proven false.
| Claim | Seen in | What Meta's sources say | Verdict |
|---|---|---|---|
| Andromeda replaced retrieval and ranking with one transformer | A vendor explainer (May 2026) | Still separate stages: in Oct 2026 Meta said retrieval brings "near ranking-level judgment", but "ranking models make the final selection". Only early-stage ranking was merged in (Oct 2025). | Not supported |
| Hook rate is the first signal Andromeda reads | A creative-tool blog (May 2026) | No Meta source on Andromeda mentions hook rate, and a new ad has none. A Meta Blueprint session lists hook type, a property of the ad, as a creative signal. | Not supported |
| Keep creative similarity under 40%; above 60% gets suppressed | A vendor explainer (updated Nov 2026) | Meta has published no similarity number. Its own column shows only Low, Medium or High. | Not supported |
| Meta groups ads with similar visuals and themes, and shares what it learns about delivery across those ads | Many agency and tool explainers | Meta said so in December 2025: look-alike ads are "variations of the same creative", so "learnings and delivery optimizations are shared". It credited "our ads system" rather than Andromeda by name. | Supported |
| Similar ads share one Entity ID and get one ticket to the auction | A vendor guide (Apr 2026) | Grouping is supported. "Entity ID" appears in no Meta source we found, and Meta hasn't said only one look-alike can be shown. | Unknown |
| Meta now requires at least 20 new ads a week | A tool blog (Nov 2025) | The Meta post it cites gives no number. The 20 likely traces to Meta's 2024 advice to keep 20 diversified ads live, which isn't a weekly quota. | Not supported |
| The Andromeda update finished rolling out in October 2025 | The same vendor explainer | The engineering post is dated December 2024, and Meta described a rebuilt Andromeda on Oct 28, 2026. | Not supported |
New concept or variation?
Every ad mixes four things: the persona it's for, the angle (the problem it solves or the reason to buy), the format and the hook. A concept is a persona plus an angle. Format and hook are how you make it, so one concept is usually a group of ads.
| What you change | What you get |
|---|---|
| The persona (who it's for, and who's on screen) | A new concept |
| The angle (the problem, the claim, the reason to buy) | A new concept |
| The format, same idea (UGC video, static, carousel) | A new execution of the same concept. In our experience it only reads as different if the first frame looks different. |
| The hook, same footage | An iteration |
| Color, background, headline, CTA or crop | An iteration. Meta's own example is identical visuals with a different text CTA. |
The fourfive case study describes the usual trap: when something won, "the team doubled down on it and creative diversity fell away".
How to tell whether a concept is working
Compare a concept only with ads that faced the same audience, auction and attribution rules. Three things trip teams up here, and in our experience general AI models make the same mistakes.
Judge a concept against its neighbors. Roll results up by concept, then compare each ad with the other ads in the same ad set over the same dates. An ad in another ad set or another month faced a different audience and auction, so the comparison says little about the creative.
Don't read the account as prospecting and retargeting. Advantage+ sales campaigns reach new, engaged and existing customers from one campaign, so a cheap CPA often says more about who an ad reached than about the concept. Use audience segments to split the reporting.
Account for March 2026. On March 3, 2026, Meta announced that click-through attribution would count link clicks only, moving shares, saves and other non-link clicks to engage-through attribution. A concept that seemed to slip in spring may just be counted differently, so compare periods on the same side of the change.
A quick audit of your live ads
- Export the last 30 days of ads with spend, plus each ad's first frame or thumbnail.
- Tag each ad with its persona, angle, format and hook. If you can't name the persona or the angle, write "none". That's a finding too.
- Count concepts, not ads. Forty ads across three concepts is three bets.
- Check where the money went. If your biggest concept takes most of the spend, the account depends on one idea, however many ads you upload.
- Run the thumbnail test. Put first frames side by side at phone size with the sound off. Any you can't tell apart count as one. If your account has the Creative diversity column, read it per ad set as a hint, not a score.
The output is a short list of personas and angles you aren't running. Those are next week's new concepts.
How many concepts a week?
Meta hasn't published a number. The biggest public benchmark we've found is Motion's Creative Benchmarks 2026. Motion says it covers over $1 billion of Meta spend across about 6,000 accounts. Its tier summary puts the median at about seven new creatives a week for accounts in its $50k to $200k spend tier, and about 19 at $1M and up (checked November 18, 2026).
Those figures count creatives, not concepts. If six of your seven ads this week are versions of one idea, Meta's December 2025 description suggests the system sees something closer to two creatives than seven. So set the split before anyone writes a brief. At seven a week, that might be two new concepts with two executions each, plus three iterations on your current winner.
Briefing checklist
Before a brief goes to a designer or creator, check that it:
- names the persona and the angle, and says whether each is new to the account
- says what's different from the live ads (a new format or hook on its own is an iteration)
- describes the opening frame, so it won't look like ads you already run
- shows the evidence for the angle, such as a customer's own words or a past result
- gives three to five hooks as separate cuts on the same body
- applies your do-not-say list before production, not in review
- labels concept or iteration in the ad name, so results roll up by concept
Where Yolodex fits
The hard part is doing all this every week: tagging every ad, reading each concept against its neighbors and keeping up as Meta changes. Yolodex was built by an ex-Meta AI team specifically for how Meta delivers ads after Andromeda, and it does this account-reading off the shelf. We tailor it to your brand at the start, so it's useful from the first week. At Grüns, a lean team of four creative strategists briefs six brands with Yolodex.
Questions
What is Meta Andromeda?
Andromeda is Meta's ad retrieval system, the first stage of ad delivery. For each person it narrows tens of millions of candidate ads down to a few thousand, and ranking models then choose which of those to show. Meta described it in December 2024 and a rebuilt version in October 2026.
Does Andromeda penalize similar ads?
Meta hasn't said it penalizes them. It said in December 2025 that its ads system groups ads that share key visual and thematic attributes and treats them as variations of the same creative, sharing what it learns about delivery across them. So ten near-copies give the system fewer distinct options than ten genuinely different ads.
How many new ad concepts a week does Meta Andromeda need?
Meta hasn't published a number. In November 2024 it advised keeping at least 20 diversified ads in Advantage+ shopping campaigns (now Advantage+ sales campaigns) and said its best performers launch new ads every week. Motion's Creative Benchmarks 2026 puts the median at 6.7 new creatives a week for accounts in its $50k to $200k spend tier, and those are creatives, not concepts.
Is a new hook a new concept?
Not on its own. A new opening on the same footage is an iteration: Meta's own example of iteration is two ads with identical visuals and different text CTAs. It becomes a new concept when the persona or the reason to buy changes.