TL;DR
Meta Andromeda is Meta's AI retrieval engine. It decides which ads even reach Meta's ranking and auction by selecting a few thousand relevant candidates from tens of millions of potential ads for each impression. You can't switch it on or off. Alongside Meta's wider automation, it tends to shift performance from targeting toward creative: distinct concepts, cleaner conversion signals and simpler structures likely matter more, while hard audience controls still apply.
Definition: Meta Andromeda
Meta's personalized ads retrieval system, first described by Meta Engineering on December 2, 2024. It selects "a few thousand relevant ad candidates" from "tens of millions of ad candidates." Larger downstream ranking models, and the auction, then determine which ads are shown.
The Three Gates (AdAdvisor's simplified advertiser model, not Meta's published stage map). Before an ad earns an impression, it helps to think of three gates:
The Three Gates: AdAdvisor's simplified advertiser model
| Gate | What happens | Meta system |
|---|---|---|
| 1. Eligibility | Policy, objective, audience controls, budget and placements determine whether an ad can be delivered for this opportunity | Campaign settings and ad review |
| 2. Retrieval | Andromeda selects a few thousand relevant candidates from tens of millions of potential ad candidates | Andromeda |
| 3. Ranking and auction selection | Downstream ranking and auction processes use predicted value, including bid, estimated action rate and ad quality, to determine what is shown | GEM-informed ranking models, Lattice architecture, the auction |
After the gates, one question sits outside Meta entirely: were the resulting conversions profitable at your margins? That part is yours.
This guide is for DTC brands, Shopify stores and media buyers who run Meta ads and want to know what Andromeda actually changed, based on Meta's own engineering and help documentation. Throughout, we separate three levels of confidence: what Meta documents, what we think it means in practice (labelled AdAdvisor operating hypothesis), and what our own product does.
Drowning in Meta Ads?
Put your campaign on autopilot with Nova.
Read moreAndromeda timeline: what Meta has said, and when
Andromeda is back-end delivery infrastructure that Meta introduced gradually, not an advertiser-facing feature rollout. Meta hasn't published every surface, country or traffic share, but its own statements give a clear sequence:
Andromeda timeline, from Meta's own statements
| When | What Meta said | Source |
|---|---|---|
| Second half of 2024 | Meta "began introducing" the Andromeda model architecture to power ads retrieval | Q2 2025 earnings call |
| December 2, 2024 | Meta Engineering described Andromeda publicly, with +6% retrieval recall, +8% ads quality on selected segments and a 10,000x increase in model capacity | Meta Engineering |
| April 2025 | Lattice, first deployed in later-stage ranking in 2023, began expanding into earlier-stage ranking | Q2 2025 earnings call |
| Q2 2025 | Andromeda enhancements (more personalized candidate selection plus expanded Facebook Reels coverage) drove nearly 4% higher conversions on Facebook mobile Feed and Reels | Q2 2025 earnings call |
| Q2 2025 | Separately, GEM improvements drove about 5% higher ad conversions on Instagram and 3% on Facebook Feed and Reels, and recent Lattice deployments drove a nearly 4% increase across Facebook Feed and Reels | Q2 2025 earnings call |
| November 10, 2025 | Meta Engineering published a technical account of GEM as the foundation model behind its ads ranking | Meta Engineering, GEM |
Sources: Meta Q2 2025 earnings call, Meta Engineering, Meta Engineering, GEM.
How to read these numbers
Retrieval recall (+6%) measures whether retrieval keeps the ads a later system judges relevant; it is not a conversion lift. Ads quality (+8%) applies to selected segments, and Meta hasn't published enough methodology to turn it into an expected CPA or ROAS change. Model capacity (10,000x) is a complexity figure, not a claim that targeting got 10,000 times better. The conversion lifts (nearly 4%, 5%, 3%) are Meta-reported, platform-level results tied to named systems and surfaces. Don't add them together, since the changes overlap in time, traffic and measurement, and don't treat any of them as an account-level forecast.
At the same time, Meta changed its targeting tools and pushed Advantage+ as the default setup path. Advertisers felt all of these changes together, which is why "Andromeda" became shorthand for everything that changed. Our Meta Ads Updates 2026 roundup tracks the news. This article explains the mechanism and separates the changes.
What is Meta Andromeda, and why did Meta build it?
Meta Andromeda is back-end delivery infrastructure, not a campaign type or a setting you switch on. Meta Engineering's December 2, 2024 post is titled "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine" and describes retrieval as "the first step in our multi-stage ads recommendation system" (Meta Engineering). Meta describes it as a system co-designed across ML, software and hardware, built on NVIDIA Grace Hopper Superchips, and says it plans to integrate Andromeda with its own MTIA chips.
The same post points to why a new retrieval engine was needed. More than a million advertisers used Meta's generative AI tools to create more than 15 million ads in a single month. When the candidate pool grows that fast, the stage that picks candidates becomes the bottleneck, and Andromeda is Meta's rebuild of that stage.
How Andromeda fits into Meta's AI ad ranking system:
- Andromeda powers the retrieval stage, selecting candidates.
- GEM (Generative Ads Recommendation Model) is Meta's ads foundation model. It learns from ad content and engagement data, then transfers knowledge to downstream ads models. Meta places it in the ranking stage after retrieval. Despite the name, it is a recommendation model, not a tool for generating ad creative.
- Lattice is a ranking architecture that generalizes learning across objectives and surfaces in place of many smaller models (Meta for Business).
- Advantage+ is the advertiser-facing automation suite (Sales, App and Leads campaigns, plus single-step automation for audience, placements, budget and creative). It runs on top of this stack and is not the same thing as Andromeda. The Meta Advantage+ guide covers each campaign type.
Retrieval-first vs ranking-first: the key change in Meta Andromeda ads
The biggest change is where your ad can lose. Before Andromeda, advertisers mostly thought about delivery as targeting plus the auction. Retrieval now makes the first cut: for any given impression, only a few thousand of tens of millions of potential candidates move on to ranking.
ALL POTENTIAL ADS
v
GATE 1 . ELIGIBILITY
policy, objective, audience controls, budget, placements
v tens of millions of potential ad candidates
GATE 2 . RETRIEVAL . Andromeda
"Is this ad a relevant candidate for this person, right now?"
v a few thousand candidates
GATE 3 . RANKING AND AUCTION SELECTION
GEM-informed ranking models, Lattice architecture;
predicted value incl. bid, estimated action rate, ad quality
v
ONE IMPRESSION
v
OUTSIDE META . BUSINESS EVALUATION
was that conversion profitable at your margins?An advertiser-oriented simplification by AdAdvisor, not Meta's published stage map. Based on Meta Engineering (December 2, 2024) and Meta's Q2 2025 earnings call. "Gates" is our label.
Meta's advertiser documentation says the auction weighs bid, estimated action rate and ad quality (Meta, About ad auctions). That is a high-level explanation, not a full disclosure of the underlying models, so don't treat it as an exact formula for diagnosing a single impression. Meta also doesn't publish a stage-by-stage map of which advertiser inputs affect which model. Read the table below as a practical guide rather than documentation.
Advertiser inputs that may influence each gate (practical guide, not Meta documentation)
| Gate | Advertiser inputs that may influence it | Common mistake |
|---|---|---|
| 1. Eligibility | Objective, audience controls, placements, budget, policy compliance, product availability | Leaving a hard business limit as a soft suggestion, or the reverse |
| 2. Retrieval (Andromeda) | Creative content, the size and variety of your candidate pool | Assuming more near-identical variants means more chances |
| 3. Ranking and auction selection | Optimization event, conversion data, bid strategy, ad quality | Changing bids to fix every delivery problem |
Two things Meta hasn't published matter here. First, Meta's auction guidance documents bid as an auction input but doesn't say how bidding affects Andromeda retrieval, so don't assume a bid change can, or cannot, change retrieval; test bid changes as delivery interventions and judge them incrementally. Second, Meta hasn't published how retrieval treats near-duplicate ads. AdAdvisor operating hypothesis: near-duplicate variants likely add limited learning value, and testing genuinely different concepts is a better use of production time.
What actually drives selection now?
Meta doesn't disclose Andromeda's full feature set, but it describes retrieval as matching people and ads at far higher complexity than before. The Andromeda post describes higher-order interactions between people and ads, advanced interaction features, ad embeddings and hierarchical indexing. Elsewhere in the stack, Meta documents that GEM learns from ad content and from long sequences of a person's ad and organic interactions. Creative is therefore an important advertiser-controlled input, used alongside user, ad, objective and measurement signals. Meta doesn't publish a weighting that ranks creative against those other inputs.
That is where the popular line "creative is targeting" comes from. It works as a metaphor, but it isn't a technical fact. Creative is one input among several, and eligibility, the optimization event, audience controls, bids and the auction all still matter.
Targeting inputs still work, but in Advantage+ audience their role depends on whether they are controls or suggestions (Meta, About Advantage+ audience):
Advantage+ audience: controls vs suggestions
| Setting type | Examples | How Meta treats it |
|---|---|---|
| Audience controls | Locations, minimum age, languages, Custom Audience exclusions | Hard limits Meta must respect |
| Suggestions | Age range, gender, detailed targeting, Custom Audience and Lookalike inclusions | Prioritized first, but delivery can reach people outside them |
In some setups, a "further limit the reach of your ads" option lets you turn certain suggestions into controls. Practitioner Jon Loomer reports that for conversion-focused performance goals, detailed targeting and Lookalikes still can't be fully restricted this way (Jon Loomer Digital). So don't assume an interest selection is a hard boundary unless you've checked how your setup treats it. Audience settings are also not a legal-compliance tool: validate regulated-category and jurisdiction requirements separately.
Two 2025 targeting changes that are not Andromeda
These often get blamed on Andromeda, but they are separate targeting-product changes (Meta, Updates to detailed targeting):
- Detailed-targeting exclusions were removed in Ads Manager on March 31, 2025, and in boosted posts on June 10, 2025. Custom Audience exclusions are still available and are listed as audience controls in Advantage+ audience. Meta reports that in an advertiser test, median cost per conversion was 22.6% lower without detailed-targeting exclusions. Meta doesn't publish the test design, population or confidence intervals, so treat that as directional platform evidence, not an expected account-level lift.
- Detailed-targeting options were consolidated from June 23, 2025. Campaigns using affected options stopped delivering on January 15, 2026 if they were not edited.
What Andromeda changes for advertisers (and what it doesn't)
Andromeda is one part of a broader shift toward automated delivery. Meta's materials describe larger candidate pools and more personalized modeling, and Meta's help guidance encourages automation and warns against unnecessary fragmentation. Because Andromeda is built to retrieve from a larger and more varied pool, testing differentiated creative is a reasonable hypothesis. Meta hasn't published a rule linking any level of concept diversity to Andromeda performance, and none of this changes the basic economics of the auction or removes the need for human judgment about margins.
AdAdvisor operating hypothesis: the table below is our reading of how practice likely shifts. Validate each change through controlled testing in your own account.
What Andromeda changes for advertisers (AdAdvisor operating hypothesis)
| Area | Pre-Andromeda habit | What likely works better now | What did not change |
|---|---|---|---|
| Targeting | Stacked interests, many exclusions | Broad or Advantage+ audience, with controls for non-negotiable limits (location, minimum age, language, Custom Audience exclusions) | Regulated-category and legal requirements still need separate checks |
| Structure | One ad set per interest or audience | Consolidating ad sets that share an objective, optimization event and overlapping audience, which can concentrate learning signals | Separation that reflects real differences in economics, geography, language, inventory or test design |
| Creative | Many variants of one winning ad | Several distinct concepts, each with placement-fit versions | Creative still has to sell the product |
| Bids and budget | Frequent reactive bid edits | Fewer reactive edits in many conversion campaigns, so results can be diagnosed cleanly | The right bid strategy still depends on objective, constraints and test design |
| Catalog | Whole catalog in one product set | Complete, policy-compliant feeds and product sets grouped by margin or category | Meta hasn't disclosed how feed-quality signals affect retrieval |
| Measurement | Platform ROAS as the scorecard | Contribution margin, break-even ROAS, incrementality | Attribution still needs checking |
For catalog advertisers, completeness, availability and product-set configuration affect whether products can be advertised and how catalog ads are assembled. The Meta catalog ad automation guide and the fix for getting Meta to advertise your whole catalog cover that side in depth.
How to adapt to Andromeda: an 8-step checklist
AdAdvisor operating hypothesis: these steps come from account-management practice. None is an Andromeda-specific rule published by Meta, so validate each one through testing.
- Reduce needless fragmentation. Each ad set needs enough optimization events to learn, and Meta's learning-phase guidance warns against splitting delivery too thinly (Meta, About the learning phase). Consolidate ad sets that share an objective, optimization event and overlapping audience, but only when it doesn't remove a control or test you need. If yours are stuck, see how to fix Learning Limited.
- Keep the separations that reflect real business differences. Separate campaigns still make sense for geography, language, product economics, new vs existing customers, regulated categories or a test that needs clean isolation.
- Count concepts, not assets. Build creative around distinct buyer motivations, use cases, proof types, formats and creators. In our practice, a new crop or background color rarely behaves like a new concept. Meta publishes no official number of creatives per ad set; its guidance is to keep diverse assets while avoiding excess ad volume (Meta, About managing ad volume).
- Start broad on placements where you can. Begin with broad placements when your assets, compliance requirements and measurement plan support it, and judge any placement exclusion on incremental or profit-aware evidence. This is general delivery advice, not an Andromeda requirement.
- Put real requirements into controls. In Advantage+ audience, use controls for non-negotiable limits and check whether your setup treats detailed targeting as a suggestion before relying on it as a boundary.
- Strengthen your event data. Pixel plus Conversions API can make event sharing more reliable and complete for Meta's measurement and optimization. Meta doesn't say Andromeda requires Conversions API or that it improves retrieval specifically. The Meta Conversions API guide walks through setup.
- Refresh on evidence, not a calendar. Treat frequency, CTR, CPM, conversion rate, cost per result and spend concentration per concept as fatigue indicators, not proof, since auction conditions, seasonality, landing-page changes and attribution delay can produce the same patterns. The creative fatigue guide and creative testing guide cover detection and test design. To choose between letting Meta adapt creative and running a controlled test, see Advantage+ Creative vs manual A/B testing.
- Measure on profit. Judge changes on contribution margin, break-even ROAS, new-customer mix and incrementality where you can. Meta-reported lifts are platform averages, not a promise for your account.
Expert take: where the leverage likely moved
AdAdvisor's interpretation: Meta's expanded automation may make creative strategy, conversion measurement and commercial guardrails relatively more important than fine-grained manual segmentation in many accounts. That is the same shift from manual tuning to decision quality described in how AI media buying works.
A heuristic we use in account reviews is the Concept-to-Ad-Set Check. Count the materially different creative concepts and the active ad sets serving the same objective. If there are more audience splits than creative ideas, check whether the structure is fragmenting the conversion signal without adding a real business distinction. This is an AdAdvisor account-review heuristic, not a Meta benchmark and not a rule every account must meet.
This view comes from a team with more than 8 years in media buying, over $60M in managed ad spend, and an ex-Meta engineer who has built ad products. One practitioner's view that points the same way is Andrew Foxwell's (Foxwell Digital) case for creative diversity in Meta ads. Treat any exact creative quota you see online, including practitioner ones, as a hypothesis to test rather than a Meta rule.
Where an AI media buyer and creative AI fit after Andromeda
Andromeda raises the importance of supplying useful ad candidates at scale. Separately, advertisers still need a business-level way to decide what is profitable enough to fund. Those are two different problems, and they are the two AdAdvisor is built to solve. Nova is a profit-first, approval-first AI media buyer that runs your Meta ads 24/7 inside the guardrails you set. By default nothing executes until you approve it, and it uses your break-even ROAS as a guardrail in its scaling decisions. Iris, its creative specialist, renders ad creative from your real products and brand. It is built for DTC brands and Shopify stores, and it is invite-only while its founding cohort onboards. The Nova explainer covers what it can do and how the approval flow works.
Neither replaces Andromeda. Standard Meta reporting doesn't automatically know your product costs, shipping, fees, discounts and returns unless you model that value data and send it through your own measurement stack, which is the gap a margin-aware operator fills. There are other realistic routes too: Meta's native tools (Advantage+ campaigns, audience and creative), Meta-focused platforms such as Madgicx (see AdAdvisor vs Madgicx), enterprise creative automation such as Smartly, specialist agencies, and in-house teams. Whichever you choose, profit-first decisions still depend on reliable cost, order and return data. Per the AdAdvisor pricing page as of September 2026, there is a free tier, MCP-only plans from $19.99/mo, and Nova at $199/mo per business ($75/mo for the Founding 100).
Frequently asked questions
Summary
Meta Andromeda is the retrieval gate at the front of Meta's delivery system. It decides which ads become candidates before ranking and auction selection ever see them. That doesn't make targeting or media buyers obsolete, but it likely moves the leverage. Accounts that reduce needless fragmentation, count creative concepts instead of assets, keep their conversion data reliable and judge results on profit are likely to get more out of Meta's automation than accounts still tuning interests and bids.
Sources
- Meta Engineering, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," December 2, 2024.
- Meta Platforms, Q2 2025 earnings call transcript, July 30, 2025.
- Meta Engineering, "Meta's Generative Ads Model (GEM): The central brain accelerating ads recommendation AI innovation," November 10, 2025.
- Meta for Business, "AI innovation in Meta's ads ranking driving advertiser performance," March 27, 2025.
- Meta Business Help Center, "About ad auctions."
- Meta Business Help Center, "About Advantage+ audience."
- Meta Business Help Center, "Updates to detailed targeting."
- Meta Business Help Center, "About the learning phase."
- Meta Business Help Center, "About managing ad volume."
- Jon Loomer Digital, "Does Meta's Advantage+ Campaign Setup Impact Targeting Control?" May 2025.
Related reading
Industry News
Meta Ads Updates 2026: What Changed This Month
The running log of Meta Ads updates in 2026, covering attribution changes, Advantage+ expansion, creative delivery shifts, and new placements. Updated monthly.
Read more
Strategy & Planning
Meta Advantage+: The Complete Guide to All Four Types (And When to Use Each)
Meta Advantage+ in 2026: Advantage+ Sales Campaigns (formerly ASC), plus Advantage+ Audience, Creative, and Placements. What each does and when to use each.
Read more
Strategy & Planning
Facebook Ads Targeting: The Complete 2026 Guide
Every Facebook ads targeting option in 2026: broad, interest, custom audiences, and lookalikes, what Meta removed, and a framework for choosing the right targeting for each goal.
Read more
AI & Automation
AI Media Buying: How It Works, Explained (2026)
AI media buying is software that reads your live ad account, decides against your goals, and acts or proposes the change, on repeat. A plain-English guide to the mechanism, the AI Media Buying Loop, and how it compares to traditional media buying.
Read more
