Meta AndromedaMeta now reads your ads to decide who sees them.

What is Meta Andromeda?

Meta's engineers describe Andromeda as the retrieval engine that picks the few thousand ads worth ranking for each person from tens of millions of candidates, and report a 6% gain in recall and an 8% gain in ads quality on selected segments since its launch. Source: Engineering at Meta.

In plain terms: before Meta ranks ads for a person, it has to shortlist them. Andromeda is that shortlist. Only the ads it picks are ever weighed by the later ranking models, so an ad that Andromeda does not retrieve for someone simply never competes for them.

How do Andromeda, Lattice and GEM fit together?

Meta has described three pieces of its ad system in public, each doing a different job. Andromeda does retrieval: it shortlists a few thousand candidate ads for a person, as Meta's engineering post of 2 December 2024 explains. Lattice is the architecture behind ranking, the stage that scores the shortlisted ads before one is shown. GEM is a foundation model that sits behind both and teaches them.

On ranking: Meta describes Lattice as one ad ranking architecture that folds many prediction tasks and surfaces into larger models, and says it increased ad quality by almost 12% and ad conversions by up to 6%. Source: Meta AI blog.

On the foundation model, in a post dated 10 November 2025: Meta's engineers describe GEM, the Generative Ads Recommendation Model, as a foundation model that does not serve ads itself but passes what it learns to the retrieval and ranking models, and report that its launch brought a 5% increase in ad conversions on Instagram and 3% on Facebook Feed. Source: Engineering at Meta. A follow-up dated 13 October 2026 went further. Meta's October 2026 GEM update reports an 18.9% conversion improvement and a 12% ad click-through rate improvement across its ad surfaces and objectives. Source: Engineering at Meta. PPC Land notes that the post gives no baseline, time window or test method for those figures. Source: PPC Land.

For an advertiser the order is what matters. An ad that retrieval does not shortlist for someone is never ranked for them, so retrieval comes first. Each figure above is Meta reporting on its own systems, measured in its own way, so read them as the direction of travel rather than a result to expect in your account.

What does the 10,000x figure in Meta's Andromeda post mean?

It describes the model, not your ads. Meta's Andromeda post credits a custom neural network with sublinear inference cost for a 10,000x increase in model capacity for retrieval: a measure of how large and detailed the model can be, not of how many ads it handles. Source: Engineering at Meta.

Some summaries turn the figure into a claim that Meta now handles ten thousand times more ads, or that you should make ten thousand times more variations. Neither is in Meta's post. A bigger retrieval model can tell more finely which ads suit which person; it still needs ads that are genuinely different to choose between.

What did Andromeda change for advertisers?

It made the ad itself the strongest targeting signal. A retrieval engine that matches ads to people by what the ads contain rewards ads that are clearly about something: a specific problem, a specific person, a specific reason to buy. Broad audiences with distinct ads now often beat narrow audiences with one ad, because the ads do the sorting.

It also changed how many ads make sense. Meta's long-standing advice to run no more than six ads per ad set has gone from its guidance, as Jon Loomer reported, and his own advice is to start with a few genuinely different ads and add more only when they earn delivery. Source: Jon Loomer.

Why do similar ads compete with each other now?

When several ads say the same thing in the same way, the system has little reason to show them to different people. They end up competing for the same slice of the audience, and the variations add cost without adding reach. That is why a set of near-copies so often performs like a single ad.

Different ideas do the opposite. A demo, a customer's words and a price comparison each appeal to different people, so the system can find a separate audience for each. Our guide on how many ads to test on Meta covers what counts as different.

What kind of creative works best with Andromeda?

Ads with one clear idea each, said early. The opening should make plain who the ad is for and what it offers, because that is what the system and the viewer both read first. Mix formats too: video, static and carousel each reach people the others miss.

Creative was already the biggest lever. Nielsen Catalina Solutions, across nearly 500 consumer campaigns (2017), put the creative at 47% of the sales lift from advertising, ahead of reach, targeting and context. Source: Nielsen Catalina Solutions. Andromeda makes that lever more direct.

Do you still need audience targeting?

Less than you did. Location, age where the law requires it, and exclusions such as existing customers still matter. But detailed interest stacks narrow the pool the system can learn from, and many advertisers now start broad and let distinct ads find their own people. Test it on your account rather than assuming either way.

Our page on ad creative for online stores shows how this looks for a store in practice.

How should you structure campaigns for Andromeda?

Simpler than before. Many advertisers now run fewer campaigns and fewer ad sets, each with a broad audience and several distinct ads, rather than many narrow ad sets with one or two ads each. Fewer, larger ad sets give the system more data to learn from in each one, and distinct ads inside them give it real choices.

Keep separate campaigns where the goal or the offer is genuinely different, such as finding new customers versus bringing back past ones, or one product line versus another. Within a campaign, let the ads rather than the audience settings do the dividing.

Whatever the structure, the habit that matters is the same: retire ideas that stopped working and add new ones regularly. Our guide to creative fatigue covers when to retire an ad.

Which Andromeda terms are not Meta's own?

A lot of Andromeda advice uses terms that sound official. Two common ones, an "Entity ID" that groups similar ads and a "60% similarity threshold" above which ads count as the same, did not appear in any Meta source we could find, and nor did a date for Andromeda's rollout being complete. They may be useful guesses by the people who coined them, but treat them as theories, not Meta's rules.

The safe reading of what Meta has published is plainer: retrieval now decides more of who sees what, and it does that better when your ads differ in idea, not just in wording.

What should you do differently this month?

Audit your live ads and write each one's idea in a sentence. Merge or cut the near-copies. Add two or three genuinely new ideas, each in a different format, and give them a broad audience and a week. Then keep the habit: new ideas every week, built on what the last week taught you.

That weekly habit is exactly what our ad creative subscription delivers, every Monday.