TL;DR: Andromeda is the first step of Meta's ad delivery. It narrows tens of millions of eligible ads to the few thousand the ranking models then choose from. Meta rebuilt it in late 2024 so it can tell creatives apart in far more detail and cope with many more of them. For an advertiser that means the ad itself does more of the targeting. A set of distinct creatives in fewer ad sets beats one ad copied across many audiences.
What is Meta Andromeda?
Meta's engineering team introduced Andromeda on December 2, 2024, as its "personalized ads retrieval engine". Retrieval is the first stage of Meta's ad delivery. Meta says it selects a few thousand relevant candidates from tens of millions of ads. Larger ranking models then pick the final ads a person sees.
Andromeda runs on Facebook and Instagram. Meta designed it end to end, from the machine learning model to the hardware it runs on. Meta calls the result a step-function change.
What did it change?
Before Andromeda, Meta says, its retrieval stage could apply only limited personalisation. It relied on separate model stages and many rule-based shortcuts to cope with the volume of ads. The rebuilt stage uses one deep neural network with 10,000 times the model capacity. A hierarchical index lets it search a far larger pool of creatives within the same time limit.
Meta's own figures, from the engineering post: a 6% improvement in recall for the retrieval system, and an 8% improvement in ad quality on selected segments. Ad quality is Meta's measure of how relevant people find an ad.
Meta's business blog of March 27, 2025 gives the plain-words version. Andromeda is a personal concierge who knows more than your taste for shoes: it knows you wear red flip-flops at the beach. That is the level of detail the new retrieval stage works at.
Why Meta built it: the creative flood
The engineering post is explicit about the reason. Advantage+ automates audiences, budgets and placements, which puts more eligible ads in the system for each person. Generative tools multiply creatives on top of that. Meta's count at the time: more than a million advertisers had used its generative tools to make more than 15 million ads in a month.
Meta built Andromeda to use that growth. The business blog puts it in advertiser terms. As businesses upload more creatives "to support a diversification strategy", Andromeda powers the models that let Meta "pick the right creative". The engineering post adds that Meta expects the next version to return "a more diverse set of ad candidates".
What it means for your account
Three consequences follow from Meta's description. They are my reading of it, not Meta's instructions.
The ad does more of the targeting. The retrieval stage now matches a person to a creative in detail. An audience definition still limits who can see an ad, but inside that limit the creative decides who gets it. Two ads with different hooks in the same ad set reach different people.
A pile of similar ads gives the system nothing to choose from. Andromeda picks among creatives. If your ten ads are one idea with a different colour, they are one candidate. If they are ten angles, formats and hooks, they are ten.
Uneven delivery inside an ad set is the system working. When one ad takes most of the budget, retrieval has found the people it fits and the others have not. That is a reason to add a new angle, not to split the audience.
What to do, and what not to do
Do:
- Run creatives that differ in angle, format and first line, not in small edits. Three distinct ads beat ten near-copies.
- Keep them in fewer ad sets. Give retrieval a broad audience and a varied set of ads to match to it.
- Judge by cost per result over weeks, not by which ad spent most this week. Delivery that shifts between ads is normal.
Do not:
- Duplicate one ad across many ad sets to force reach. The same creative is the same candidate everywhere.
- Slice audiences to "help" delivery. Narrow slices shrink what the system can learn from and send ad sets back into the learning phase more often.
- Rebuild after one bad week. A significant edit restarts learning, which costs more than the week did.
The ad sheet in Adsap launches a batch of distinct creatives into one ad set in one pass, and the creative performance view shows how delivery spread across them.
Andromeda next to GEM and the ranking models
Andromeda retrieves; it does not decide. After it narrows the pool, Meta's ranking models choose the ad a person sees. Meta has described two of them. The Generative Ads Recommendation Model, GEM, rolled out across its ads system in 2025. The Adaptive Ranking Model launched on Instagram in late 2025, by Meta's account. Meta's 2026 engineering posts list Andromeda next to them as one part of the chain. One of those posts records a 60% gain in Andromeda's inference throughput from automated kernel optimisation. A faster retrieval stage considers more candidates per request, which leaves more room for creative variety to matter.
Frequently asked questions
No. It is a rebuilt retrieval stage inside Meta's ad delivery, introduced in December 2024 and running since. There is no date on which it switched on for your account and nothing to recover from. It changes what the system can tell apart, which is why creative variety matters more than it did.
Not the settings. Audiences, placements and budgets work as before. What changed is that retrieval now matches people to creatives in far more detail inside those settings, so the ad itself carries more of the targeting.
Meta publishes no number. The useful rule from Meta's description is variety, not count. Ads that differ in angle, format and first line give retrieval distinct candidates, and near-copies do not. Three distinct ads beat ten copies.
No. It is part of Meta's delivery system for every advertiser on Facebook and Instagram, and there is no setting for it in Ads Manager or the Marketing API.
Sources
All Meta pages read on 19 October 2026.
- Engineering at Meta, Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, December 2, 2024. The definition of retrieval, the 10,000 times model capacity and the hierarchical index. The 6% recall and 8% ad quality figures, the generative-tools count, and the "more diverse set of ad candidates" line.
- Meta for Business, AI Innovations in Meta's Ad Ranking Driving Advertiser Performance, March 27, 2025. The concierge analogy, the 8% ad quality figure, the creative diversification sentence, and GEM.
- Engineering at Meta, Meta Adaptive Ranking Model, March 31, 2026; KernelEvolve, April 2, 2026; and Exploring Hierarchical Interest Representation for Meta Ads, July 15, 2026. Where Andromeda sits next to GEM and the ranking models, and the 60% throughput figure.
- Adsap docs, ad sheets and the performance dashboard.