Andromeda: The End of Audience Targeting as You Know It

Most of what gets written about Meta’s Andromeda describes something bigger and vaguer than what Meta actually published. The primary source is a December 2024 post from Meta’s engineering team, and it is specific. Andromeda is a retrieval system: the first stage of Meta’s ad delivery, which narrows tens of millions of candidate ads down to a few thousand for a given person before more complex ranking models pick what is shown.

Since then Meta’s engineers have kept publishing about the stages that come after retrieval, including a multi-stage architecture for ads ranking and how they train GEM, a large foundation model used in ads recommendation, both in August 2026. Andromeda is one layer of a stack. Understanding which layer it is explains what it changes for your targeting, and what it doesn’t.

What retrieval is, and why it was the bottleneck

Every time someone opens Facebook or Instagram, Meta has to decide which ads to consider for them. It can’t run its most expensive models on every eligible ad, so delivery works in stages. Retrieval comes first and does the heaviest cut. Meta’s post says retrieval processes three orders of magnitude more ads than the stages after it, under tight time limits.

Before Andromeda, according to the same post, Meta’s retrieval applied only limited personalization, using isolated model stages and many rule-based heuristics to cope with the volume. That matters for advertisers because retrieval is a gate. An ad that isn’t retrieved for a person never reaches the auction for that person, whatever its bid or quality.

Meta also explains why the old approach was under strain. Its Advantage+ tools automate audiences, budgets, placements and creative generation, which multiplies the number of eligible ads, and generative AI tools were producing more ads still. The post reports that more than a million advertisers used Meta’s generative AI tools to create over 15 million ads in a single month. Retrieval had to handle far more candidates without slowing down.

What Andromeda changed

Meta’s description has three parts worth knowing.

A much larger model. Andromeda uses a deep neural network built for NVIDIA Grace Hopper hardware, which Meta says allowed a 10,000-fold increase in retrieval model capacity. The model learns relationships between people’s interests and the products and services in ads, replacing a patchwork of rules with one system that can be optimized end to end.

A learned index. Ads are organized in a layered index that is trained together with the retrieval model, so the system can search a huge pool efficiently by exploring only the most relevant branches. Meta says this improves both precision and recall compared with common alternatives.

Measured gains. Meta reports a 6 percent improvement in retrieval recall and an 8 percent improvement in ads quality on selected segments. Those are Meta’s figures, from Meta’s systems. The many larger percentages attributed to Andromeda in agency guides generally come from somewhere else.

What the post does not say is equally useful. It doesn’t describe a creative similarity score, a rule that merges near-identical ads, or a required number of creatives. Those claims circulate widely, including in an earlier version of this article. They aren’t in Meta’s account of the system.

What it changes for audience targeting

Here is the distinction most coverage blurs. Andromeda didn’t delete audience targeting. Your targeting settings still define which ads are eligible for which people. What changed is how much work those settings do.

When retrieval relied on rules and limited personalization, the advertiser’s audience definition did much of the matching. With a large model trained on how people respond to ads, the model does the matching, and a narrow manual audience mostly restricts the pool it is allowed to choose from. Your targeting shifts from instruction to constraint. Useful constraints remain useful: excluding existing customers, respecting a service area, keeping to a legal age limit. Guessing interests for the model usually isn’t, because the model is better placed to find buyers than a hand-built audience.

That is why broad audiences and consolidated campaigns tend to suit Meta’s current delivery. A campaign split into many small, narrowly targeted ad sets gives each part less data and less room, the structural cost laid out in how to lower ad costs.

Why the creative matters more

If the model does the matching, the ads are what it matches with. Meta’s post frames Andromeda as built for a world with far more ads in the system, and describes future work aimed at retrieving a more diverse set of ad candidates. An account that gives the model several genuinely different ads gives it several different ways to find a fit with different people. An account that gives it many near-copies of one idea gives it far less to work with than the ad count suggests.

That practical consequence, what kind of creative variety the system needs and what Meta’s own tools can produce for you, is covered in what Meta’s ad automation needs from your creative.

Two things stay firmly with you. The first is the signal: the conversion events you send teach the whole stack what a good outcome looks like, and bad signal teaches it to retrieve the wrong people efficiently. The second is the creative. Setting up accounts so both are right, and letting the machine do the matching, is how performance marketing built around Meta’s current delivery differs from managing audiences by hand.

Andromeda did not end targeting. It moved targeting from the settings you choose into the model Meta trains, and left you in charge of the two things that model learns from.

Frequently Asked Questions

What is Meta Andromeda?

Andromeda is the retrieval system in Meta’s ad delivery, described by Meta’s engineering team in December 2024. It is the first stage of choosing which ads a person might see, narrowing tens of millions of candidate ads to a few thousand before ranking models decide which to show. It replaced a more rule-based retrieval process with a much larger neural network and a learned index.

Did Andromeda get rid of audience targeting on Meta?

No. Targeting settings still determine which ads are eligible for which people. What changed is that Meta’s models do more of the matching within that eligible pool, so narrow interest audiences mainly limit the model’s options. Constraints that reflect real business rules, such as exclusions, locations and age limits, remain useful. Guessing interests on the model’s behalf usually adds less than it used to.

What results has Meta reported from Andromeda?

In its engineering post, Meta reported a 6 percent improvement in retrieval recall and an 8 percent improvement in ads quality on selected segments, along with a 10,000-fold increase in retrieval model capacity. Larger performance figures attributed to Andromeda in marketing guides are generally advertiser or vendor observations, not numbers Meta published about the system. Treat them as claims to test in your own account.

How is Andromeda different from GEM?

They sit at different stages. Andromeda handles retrieval, deciding which few thousand ads are considered for a person. GEM is a large foundation model Meta uses in ads recommendation, and Meta’s engineers have written about it separately, including an August 2026 post on how they train it. Retrieval decides which ads get considered. The later ranking stages decide which one is shown.