Andromeda is the system Meta uses to pick, from tens of millions of ads, the few thousand that qualify for one person, before the auction decides the winner. The broader your audience, the more your ad creative determines who your ad reaches. That’s why, in Meta Ads, your creatives have become the most important lever for your targeting, and your audience settings less so.
That leads to choices about your account structure, your creatives and your patience: few ad sets, many different angles and hands off during the learning phase.
What Andromeda is
Meta described Andromeda in December 2024 on its engineering blog. Showing an ad at Meta happens in steps:
- Retrieval: choosing a shortlist of a few thousand candidates from tens of millions of ads. That’s the Andromeda step.
- Ranking: deciding from that shortlist which ad someone gets to see. For this, Meta uses GEM, its generative ads model, among other things.
Andromeda runs on NVIDIA Grace Hopper chips. According to Meta, a new neural network gives the retrieval models 10,000 times more capacity on that hardware. Meta wants to scale this further with MTIA, its own AI chips. Meta itself cites 6% better recall and 8% higher ad quality on selected segments. These are Meta’s figures about its own system.
What changes for advertisers
Meta built Andromeda to work with many more ads. The reason Meta gives is that advertisers are supplying more and more creatives because of Advantage+ and generative AI. The system chooses per person which ad fits, based on signals about that person and about the ad itself.
Two things follow from that, and they seem to contradict each other:
- Variety pays off. For Advantage+ sales campaigns, Meta recommends supplying a wide range of different ad assets, so relevance per person goes up. With one angle, you reach one type of buyer.
- Volume costs learning. Meta also writes that the system learns less about each ad and ad set if you create a lot of them, and recommends combining similar ad sets.
They fit together once you separate variety from volume. Variety means different angles within a few ad sets. Volume is a pile of nearly identical variants, spread across many small ad sets.
Why granular accounts will struggle
According to Meta, an ad set usually exits the learning phase after about 50 results in the week after the last significant edit. If a set doesn’t reach that, it gets the Learning limited status. Among the causes, Meta lists a small audience, a low budget, auction overlap and too many ads at once.
An account with ten ad sets on ten interest groups spreads its conversions across ten learning paths. Each path then gets too few results to learn. Meta’s first fix for Learning limited is to combine ad sets and campaigns. So a structure that helped the old system works against the new one.
Decision rules for your Meta account
- Number of ad sets: as many as your budget can give each set about 50 results per week. If you don’t reach that, I merge sets until you do.
- Optimization event: Meta suggests optimizing for add to cart when you have too few purchases. For a webshop, I only do that if merging isn’t enough, and I go back to purchases as soon as volume allows.
- New creative: I count a creative as new if the angle is different: a different opening line, a different format (video, static, carousel), a different message or a different buyer in the picture. Ten color variants of one image are one angle.
- Audience: Advantage+ audience is my starting point. Hard limits, such as a minimum age or countries you don’t ship to, go in the account settings. I exclude existing customers with a custom audience when the goal is new customers.
- Intervening: not during the learning phase. Meta recommends waiting to edit until a set has exited the learning phase, because every edit restarts learning. I evaluate per week, the window Meta itself uses for the learning phase.
- Tracking: Andromeda and the ranking learn from the conversions you send back. The pixel and the Conversions API both belong in your setup, ideally via server-side tracking, with deduplication so a purchase doesn’t count twice.
Your role shifts from building audiences to coming up with angles. You provide the positioning, the messages and the boundaries, and the system makes the call per auction. Which controls you still have in your own hands in Advantage+ is covered in my blog post on Advantage+.
My advice
See your creatives as your most important optimization lever. Build a fixed rhythm in which you regularly add new, clearly different ads to a small number of ad sets. And give the system the time and volume to learn before you draw conclusions.
Want to know whether your Meta account is still built the old way? Request a Free Performance Quick Scan, and I’ll look at your structure, your creatives and your tracking.
Sources
- Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine (Engineering at Meta, December 2, 2024)
- AI Innovation in Meta’s Ads Ranking Driving Advertiser Performance (Meta for Business, March 27, 2025)
- Meta’s Generative Ads Model (GEM) (Engineering at Meta, November 10, 2025)
- About the learning phase (Meta Business Help Center)
- About Learning limited (Meta Business Help Center)
- About Advantage+ sales campaigns (Meta Business Help Center)
- What causes Advantage+ to turn on and off (Meta Business Help Center)