Insights·November 12, 2026

Meta Andromeda: What It Is and What It Isn't

Meta Andromeda is the retrieval stage of Meta's ad system, not the algorithm itself. What it does, what changed in 2026 and what the entity ID myth gets wrong.

Christoph Schütte

Meta Andromeda is Meta's ad retrieval engine, announced in December 2024. For every ad request it narrows tens of millions of eligible ads to a few thousand candidates, and ranking models and the auction then decide which ad a person sees. It is one stage of Meta's ad system, not the algorithm itself, and Meta rebuilt it in 2026.

Key takeaways

• Andromeda shortlists a few thousand ads out of tens of millions; models like GEM then choose and rank, and the auction lets only one ad per advertiser compete.
• Meta's 2026 version scores every eligible ad at over one billion ads per second per GPU server, 100 times the 2024 system, and Meta credits Andromeda with 2.4% more conversions on Facebook since 2025.
• The popular "entity ID" explanation comes from an unpublished Meta partner deck about shared learnings, not from anything Meta has published about Andromeda.
• Meta does group look-alike ads, but they share learnings and delivery; they do not collapse into one auction ticket, and your own ads never bid against each other.
• In December 2025 Meta replaced its only creative count, at least 20 ads, with "a broad portfolio," and its help center advises fewer ads per ad set with more diverse assets.

What is Meta Andromeda?

Andromeda is the bouncer, not the judge: the first stage of Meta's ad delivery, not the whole delivery engine. In its Q4 2024 earnings call, Meta described ads retrieval as the part of the ranking process "where we narrow down a pool of tens of millions of ads to the few thousand we consider showing someone." Larger ranking models then score that shortlist, and the auction decides what wins.

The split is economic. Retrieval must judge tens of millions of ads in milliseconds, so each judgement has to be cheap, while ranking can afford a far heavier model for a few thousand. An ad that is never retrieved cannot win, but being retrieved is not being chosen.

Meta's engineering post from December 2, 2024 describes the first retrieval system: a deep neural network built for the NVIDIA Grace Hopper Superchip and Meta's MTIA chips, with a hierarchical index trained together with the model instead of the common two-tower setup, which Meta says raised retrieval model capacity 10,000 times. A Meta paper from August 2024, Hierarchical Structured Neural Network, describes the same approach and shares four authors with the 2026 Andromeda post, though Meta has never called them identical.

Then Meta rebuilt it. In an engineering post from October 28, 2026, Meta describes the next generation: a retrieval mixture-of-experts model that scores every eligible ad "with near ranking-level model performance at retrieval scale," at over one billion ads per second on a single GPU server, a 100x lift over the 2024 system. The same post says that since 2025, Andromeda "has delivered a 2.4% increase in conversions on Facebook."

Andromeda can only shortlist the ads you give it, for Facebook ads and Instagram ads alike, which is why our creative testing workflow starts with what goes into the pool.

Before you accept a claim about "the Meta algorithm," ask which stage it is about. Most claims that sound contradictory describe different stages.

What changed in the Meta Andromeda update?

There was no single update. Meta changed its retrieval system several times, and the dated changes sit in its earnings calls, not its blog posts. In the Q3 2025 earnings call, Meta said it had "significantly improved performance of Andromeda in Q3 by combining models across retrieval and early stage ranking into a single model, driving a 14% increase in ads quality on Facebook surfaces." That is the change advertisers felt in autumn 2025. The "October 2025 global rollout" that many posts cite has no Meta announcement behind it; the earliest mention we traced is an uncited agency post from October 16, 2025.

In the Q4 2025 call, Meta said Andromeda now runs on NVIDIA, AMD and its own MTIA chips, which, along with model changes, "enabled us to nearly triple Andromeda's compute efficiency." It is not mentioned in the Q1 or Q2 2026 calls; in the Q2 2026 call, Meta introduced Meta Generative Recommender, which uses large language models "to reason about ad content and user preferences together." The October 2026 rebuild shows that Andromeda was never retired; Meta simply stopped talking about it on investor calls.

Part of the confusion is the name itself. Ads consultant Jon Loomer put it well: "If Meta Andromeda were simply labeled 'The Ad Retrieval Layer,' I'm convinced we wouldn't hear a word about this." The other part is how Meta's numbers travel. The same 8% gain appears three ways:

The version without a qualifier is the one most often copied, and none of the three says what "ads quality" measures. None of that makes the number false. It makes it narrower than it sounds.

Does Andromeda cluster your ads by entity ID?

No. When we checked in August 2026, Google's AI Overview for "meta andromeda" said Andromeda "clusters similar ad creatives using semantic indexing so distinct creative variations act as separate tickets to the auction." The term "entity ID" does not appear in Meta's Andromeda posts or in any Meta publication we checked.

It is not invented, though. It comes from a Meta partner-training deck, "Growth Academy: Creative Diversification," shown in a media buyer's Instagram reel on February 20, 2025, which says ads with the same Entity ID are treated "as the same entity, sharing the learnings and delivery." That is a claim about delivery, not retrieval. In late September and October 2025, agency posts, a widely watched video and ad-tool blogs moved the term into "Andromeda's hierarchical index," and one guide added a 60% similarity threshold Meta has never published.

What Meta did publish came on December 16, 2025, in Demystifying Creative Diversification:

Our ads system also analyzes creatives and groups together ads that share key visual and thematic attributes. When multiple ads look or feel alike, these are seen as variations of the same creative, meaning learnings and delivery optimizations are shared at the creative level – not just the individual ad level.

So the grouping is real, and the folklore gets two things wrong. Wrong stage: the same post describes Andromeda separately, as the step that "scans tens of millions of ads and selects a few thousand strong candidates," after which "models like GEM choose and rank." Grouping is a property of the ads system, not of the retrieval index. Wrong consequence: grouped ads share learnings and delivery optimizations. They do not collapse into a single ticket to the auction.

Why do near-identical ads still hurt performance?

Because of two mechanisms Meta documents in its help center, and neither is Andromeda. First, your ads never compete with each other: according to Meta's page on auction overlap, when several ads from one advertiser enter the same auction, only the one with the highest total value is considered. Overlap "can lead to limited delivery when ads from the same advertiser share the same audience or assets," and it can stop an ad set from spending its budget or leaving the learning phase.

Second, volume dilutes learning. Meta's page on managing ad volume says that when an advertiser runs too many ads at once, each ad delivers less often and fewer ads exit learning. Its advice: decrease ads per ad set but keep diverse creative assets, since one ad can hold up to 10 of them. An ad set usually needs about 50 results in the week after its last significant edit to leave the learning phase.

Meta's guidance on creative diversity also separates two problems that get mixed up: creative fatigue is overexposure of the same creative, creative similarity is assets that look too alike, even as different files. Matt Steiner, a Meta vice president who works on ads ranking, made the practical point in a June 2026 interview with Sam Piliero: an extremely large number of ads probably won't help more than a reasonable diversity of them. One team we work with saw the same thing from the other side: five hooks on one video stopped paying off, while three hooks still helped when a new concept was first tested.

How many ads should you run per ad set?

Meta no longer gives a number. In November 2024, it recommended at least 20 diversified ads in Advantage+ shopping campaigns and 20 to 30% of the budget for testing new ads. In December 2025, it told advertisers that "instead of producing a fixed volume of ad creatives per campaign," they should build "a broad portfolio of assets" across text, image and video, based on several customer personas.

Your budget sets the practical ceiling, because every ad set needs about 50 results a week to learn. Divide your weekly conversions by 50 and you get the number of ad sets your account can feed. A worked example: at $40,000 a month and a $50 cost per purchase, a brand buys about 800 purchases a month, roughly 185 a week, which is enough for three ad sets.

For scale, Motion's Creative Benchmarks 2026 show what accounts actually ship, not what Andromeda needs; the report never mentions Andromeda. Its methodology covers 578,750 creatives from 6,015 advertiser accounts and $1.29 billion in Meta spend between September 2025 and January 2026:

Our ad volume calculator runs this math for your spend, and our analysis of creative volume benchmarks explains why the median, not the mean, is the number to plan with.

What should you change in your ad creative strategy?

Strip out what Meta has not said, and a short list for the Andromeda era remains:

1. Plan concepts, not hooks. Meta defines true creative diversification as distinct assets "tailored to different personas or use cases," different in visuals, messaging and formats.
2. Count variants as one idea. New hooks, crops and colors on the same video are grouped and share learnings. Use them to refresh a proven concept, not to feed the system new ideas.
3. Consolidate. Combine similar ad sets, as Meta's help center advises, and run fewer ads per ad set with several assets each, so every ad set collects enough results to learn.
4. Judge after the learning phase. Edits during learning restart it, so compare concepts on cost per result once each ad set has its 50 results.
5. Treat fatigue and similarity separately. Fatigue means the same people have seen an ad too often; similarity means your ads look alike. Our guide to ad fatigue covers the first.

How to produce distinct concepts with Solid

Distinctness is a production problem, not an upload problem: fifty variants of one concept cost almost nothing, genuinely different concepts take real work, and the grouping means the cheap option is the one that does not help. Solid, the AI creative engine for brands and their agencies, attacks that constraint with the footage you already own. The Smart Library tags every scene, so a shot for a new persona is a search away; the Solid Agent turns a brief such as "three concepts for first-time buyers" into finished ads in your brand's style; and the AI Video Generator fills only the shots you never filmed. See how teams run weekly concept tests on our creative testing page.

FAQ

Is Meta Andromeda the same as Advantage+?

No. Advantage+ is Meta's suite of automated campaign tools for budgets, audiences, placements and creative. Andromeda is the retrieval engine underneath that decides which ads are considered for each person. Meta's October 2026 post says the pool of eligible ads keeps growing "driven by Advantage+ automation and generative AI creative tools," which is why retrieval had to scale. There is nothing to switch on: retrieval runs for every campaign.

What is the difference between Andromeda and GEM?

Andromeda is retrieval: which ads are considered. GEM is Meta's ads foundation model for ranking; according to Meta's GEM post, Meta trains it as a teacher whose knowledge is transferred to hundreds of user-facing models. Read its numbers closely: a "23x increase in effective training FLOPS" with "1.43x" higher utilization "using 16x more GPUs" is one gain, since 16 × 1.43 ≈ 23.

Does Andromeda still use your targeting?

Yes, as a filter. Meta's October 2026 post says Andromeda's index matching enforces advertisers' targeting rules, evaluating "complex advertiser-defined targeting logic across tens of feature categories" before ads are scored, so even with broad targeting your limits apply first. Audience targeting decides which ads are eligible for a person; scoring decides which of them make the shortlist, and that is where distinct creative matters.

Ready to feed Meta better ideas?

Turn the footage you already own into distinct concepts for every customer type, every week.

Start free trial or Book a demo

About the author

Christoph Schütte is Co-Founder and CTO of Solid, the AI creative engine for brands and their agencies. A physicist and former Engineering Lead at Google, he worked on video intelligence and computer vision in Google's self-driving cars and robots unit and with brands at Google Shopping on the loop between ad performance and production. Connect on LinkedIn

Sources

• Meta Engineering: Meta Andromeda, December 2, 2024 and October 28, 2026; GEM, November 10, 2025
• Meta for Business: creative diversification steps, November 19, 2024; AI innovations in ad ranking, March 27, 2025; Demystifying Creative Diversification, December 16, 2025
• Meta Business Help Center: auction overlap, managing ad volume and the learning phase, checked November 12, 2026
• Meta earnings call transcripts: Q4 2024, Q3 2025, Q4 2025, Q1 2026 and Q2 2026
• Rangadurai and others (Meta): Hierarchical Structured Neural Network, arXiv, August 2024
• Jon Loomer Digital: Meta Andromeda and ad retrieval
• Sam Piliero: interview with Meta's Matt Steiner, YouTube, June 4, 2026
• Motion: Creative Benchmarks 2026, methodology
• Solid research on the entity ID claim, August 2026, and customer calls (customers not named)