--- title: 'Meta Andromeda: what changed and what to do' description: 'What Meta Andromeda does, how it differs from Lattice and GEM, and which creative advice Meta actually documents. Sources checked October 22, 2026.' canonical: 'https://plainrouter.com/library/meta-andromeda' format: Guide published_at: '2026-10-22' last_updated: '2026-10-22' --- # What is Meta Andromeda, and what changed for advertisers? Meta Andromeda is the retrieval system that narrows tens of millions of eligible ads to a few thousand candidates; ranking models then evaluate that shortlist. It is one part of Meta's ad recommendation system, not the whole algorithm. Meta introduced it in 2024 and described its design publicly on December 2, 2024; later GEM and GenRecs announcements describe other developments, not a single new Andromeda launch. ## What is Andromeda in Meta ads? Andromeda helps decide which ads reach the shortlist for a particular person. Retrieval has to handle a very large candidate pool within a tight time budget, before more expensive ranking takes place. Meta's [December 2024 engineering article]() describes a jointly designed model, index and hardware system for that job. The practical distinction is between finding plausible candidates and choosing among them. A stronger retrieval system can surface a relevant creative that a weaker shortlist might miss. That does not mean every uploaded ad receives delivery, or that a new creative is guaranteed to improve sales. ## How do Andromeda, Lattice and GEM differ? Andromeda retrieves candidates, Lattice is a ranking architecture, and GEM is a foundation model whose learning improves other recommendation models. They are related parts of the system, not three names for the same update. | System | Stage | What it does | Meta-reported result | Meta source and date | | --- | --- | --- | --- | --- | | Andromeda | Retrieval | Personalizes the candidate shortlist before subsequent ranking. | Meta-reported: +6% recall improvement to the retrieval system and +8% ads quality improvement on selected segments. | [Engineering, December 2, 2024](); [Business, March 27, 2025]() repeats an 8% increase in ads quality. | | Lattice | Ranking | Generalizes learning across objectives and surfaces instead of maintaining many smaller models. | Meta-reported: ad quality increased by almost 12%; ad conversions increased by up to 6%. | [Business, March 27, 2025](). | | GEM | Foundation-model learning for ranking | Transfers learning to downstream recommendation models; the 2026 version learns across surfaces and objectives. | Meta-reported: ad conversions increased by up to 5% during the initial Reels launch. In October 2026: 18.9% conversion improvement and 12% ad click-through rate improvement across Meta's ad surfaces and objectives. | [Business, March 27, 2025](); [Engineering, October 13, 2026](). | These are Meta's reported measurements, not Plainrouter benchmarks or expected gains for your account. The March GEM result's footnote describes a randomized Facebook Reels experiment over one week in March 2025, measuring ad conversion rate normalized by impressions. The October GEM figures cover a different scope. Do not add the percentages together or treat them as one advertiser's before-and-after result. ## What changed in the Meta Andromeda update timeline? The public record spans several systems and dates. Calling all of it the “Meta Andromeda update” hides those distinctions. - December 2, 2024: Meta publishes [Andromeda's retrieval design](), including hierarchical indexing and increased model capacity. This is a publication date, not evidence of a worldwide rollout on that day. - March 27, 2025: Meta's [business explanation]() separates Andromeda, Lattice, GEM and sequence learning, and reports results for each. - November 10, 2025: Meta details [GEM's architecture and knowledge transfer](). The foundation model helps other ads models learn; it is not an advertiser-facing image generator. - January 28, 2026: [2026: AI Drives Performance]() discusses expanded GEM training, sequence learning and Lattice consolidation. This is a broader performance update, not a new Andromeda launch date. - October 13, 2026: Meta describes [GEM as a central model across surfaces and objectives](), with learning transferred to serving models. - October 19, 2026: Meta publishes [GenRecs Tokenization](). Items are represented through a shared semantic vocabulary, supporting learning across related items in retrieval and ranking. That is not a published rule that visually similar ads share one delivery limit. ## What does Meta actually recommend advertisers do? Meta recommends creative diversification: give the system meaningfully different material to match with different people. Its March 27, 2025 post connects Andromeda's increased capacity with businesses uploading a wider creative selection. Meta's [Blueprint training published May 18, 2026]() defines diversification as “varying formats, visuals and messages across ad sets”. Its [Performance 5 page](), checked October 22, 2026, says: “Develop a variety of ad creative”. Neither statement establishes a universal creative count or a public similarity threshold. For a practical test, vary the concept as well as the execution: a product demonstration, a problem-and-solution explanation, or a different use case. Name the audience, message and outcome you want to test. This is a way to turn Meta's advice into a testable hypothesis, not a Meta-prescribed formula. If results weaken, investigate [Facebook ad fatigue]() rather than assuming a model update caused it. Keep delivery changes, tracking failures and the [learning phase]() separate from your creative hypothesis. ## Which popular Andromeda claims have a Meta source? Our claim check distinguishes published engineering facts from instructions circulating in marketing commentary. “Not found” means the reviewed Meta sources do not establish the rule; it does not prove that an undocumented mechanism cannot exist. | Claim | Where it circulates | Found in a Meta source? | What Meta actually says | | --- | --- | --- | --- | | Near-duplicate ads are grouped under one Entity ID. | Agency blogs and creative-testing commentary. | Not found in Meta's published sources as of Oct 22, 2026. | The [GenRecs post, October 19, 2026]() describes related items sharing semantic tokens. It does not establish the claimed Entity ID grouping or its delivery consequences. | | A creative similarity score above X% suppresses delivery. | Agency blogs and social posts. | Not found in Meta's published sources as of Oct 22, 2026. | The reviewed [creative-diversification training, May 18, 2026]() recommends variation. It does not publish a similarity-score cutoff. | | Every ad set should run a fixed N creatives. | Agency playbooks and testing advice. | Not found in Meta's published sources as of Oct 22, 2026. | [Performance 5, checked October 22, 2026](), recommends a variety of creative. The reviewed guidance does not make a fixed count an Andromeda requirement. | | Andromeda is 100× faster or evaluates 10,000× more ads. | Agency summaries of the engineering post. | The numbers appear, but those blanket interpretations are not what the [December 2, 2024 source]() measures. | The 10,000× figure concerns model capacity; Meta's March 2025 explanation calls it model complexity. Over 100× concerns feature-extraction latency and throughput compared with previous CPU-based components. Neither is an account-wide delivery-speed guarantee or a count of ads evaluated. | | Andromeda replaced targeting. | Agency blogs and creative-strategy commentary. | Not found in Meta's published sources as of Oct 22, 2026. | Meta's [March 27, 2025 explanation]() describes personalized retrieval that supports Advantage+ automation. It does not announce the removal of advertiser targeting controls. | Method: Plainrouter read the linked Meta engineering posts, the March 2025 business explanation, Performance 5 and the Blueprint training description on October 22, 2026. This is a documentation audit, not a delivery experiment, an inspection of Meta's private models or an exhaustive search of every Meta publication. Generic circulation labels describe the claims reviewed; they are not evidence that the claims are true. ## What does this mean for agents and AI-generated creative? Meta's diversification guidance supports testing distinct concepts, not producing near-copies simply to increase volume. AI can make variations easier to prepare, but a person still needs to check the offer, factual claims, rights and destination before launch. Use the [review steps for AI-generated Meta ads](), and distinguish creative preparation from the decision to activate an ad. ## Go deeper - [Creative diversification](): review concepts, messages, visuals and formats before launch. - [Campaign and ad-set structure](): distinguish documented limits from planning choices. - [Creative-testing methods](): compare experiments, in-campaign tests and ordinary delivery. - [Facebook ad fatigue](): diagnose a decline before attributing it to creative repetition. ## Frequently asked questions ### When did Meta Andromeda launch? Meta says Andromeda was introduced in 2024, and published its engineering explanation on December 2, 2024. The reviewed sources do not establish that every account switched on one worldwide launch date. The October 2026 GEM and GenRecs posts are separate developments. ### Is Andromeda the same as Advantage+? No. Andromeda is an underlying retrieval system; Advantage+ is advertiser-facing automation that can benefit from it. You do not install Andromeda as a separate campaign tool. ### Does Andromeda mean I need more creatives? Meta recommends creative diversification, but the reviewed sources do not prescribe one fixed number for every ad set. Start with meaningfully different messages and formats that you can evaluate, rather than a quota of cosmetic variations. ### What is the difference between Andromeda, Lattice and GEM? Andromeda builds the candidate shortlist. Lattice generalizes ranking across objectives and surfaces, while GEM transfers foundation-model learning to recommendation models. Improvements to one should not automatically be attributed to another. ### Is the Meta Andromeda update an AI agent that runs my ads? No. Retrieval and ranking decide which ads to consider and show; an advertiser's agent workflow prepares or changes account objects through separate interfaces. Compare [AI agents with automated Meta rules]() if that is the task you are choosing software for. ## Sources - [Meta Engineering: Andromeda, December 2, 2024](). - [Meta for Business: AI innovations in ad ranking, March 27, 2025](). - [Meta Engineering: GEM, November 10, 2025](). - [Meta Newsroom: 2026: AI Drives Performance, January 28, 2026](). - [Meta Engineering: GEM scaling, October 13, 2026](). - [Meta Engineering: GenRecs Tokenization, October 19, 2026](). - [Meta Blueprint: creative diversification, May 18, 2026](). - [Meta Performance 5, checked October 22, 2026]().