Meta Andromeda: what changed in ad retrieval, and what advertisers do about it

Meta Andromeda is a personalized ads retrieval engine. Meta's engineering team introduced it on 2 December 2024 as the system behind the first stage of ad recommendation, which it describes as selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates. Those candidates then go to the ranking models that decide which ad is shown and at what price. For anyone asking what the Andromeda update to Meta Ads changed, that is the short answer: not who is eligible to see an ad, but which ads are in the running for each person.

The post credits Andromeda with a 10,000x increase in model capacity at the retrieval stage, achieved with a hierarchical index that scales to a large volume of ad creatives and an inference cost that grows sublinearly with it. It runs on NVIDIA Grace Hopper Superchips and Meta's own MTIA accelerators, with feature extraction over 100 times faster and end-to-end inference throughput over 3 times higher. The results Meta reports are a 6% recall improvement and an 8% ads quality improvement on selected segments.

Meta frames Andromeda as infrastructure for Advantage+, the automation that allocates budget, chooses audiences and adjusts bids. In March 2025 Meta for Business put it plainly: as businesses upload more creatives to support a diversification strategy, Andromeda powers the models that pick the right creative for each person.

Two stages decide what a person sees. Retrieval chooses which few thousand ads are even considered, and ranking scores those candidates for the auction. A targeting setting still defines who is eligible to see an ad, but Andromeda works upstream of that choice: it decides which of your ads, and which of your competitors' ads, enter the shortlist for a given person.

That is why creative has become a targeting signal. Meta's retrieval index is built to hold a large volume of creatives and to choose among them per person, so each distinct ad is a different way into the shortlist. An account with one creative per audience gives the index one chance; an account with varied angles, formats and messages gives it several. Meta's own advice to go beyond formats and vary visuals, messages and storytelling describes the same mechanism from the advertiser's side.

What Andromeda does not change is the contract after retrieval. The optimization event you choose, your budget and your bid strategy still govern how an ad set is delivered once its ads are candidates, and the learning phase still applies: Meta says performance is less stable and CPA usually worse while an ad set is in it. Meta's help pages on fragmentation, auction overlap and ad volume predate the model and still apply.

Creative diversity first. On its Advantage+ sales campaigns page Meta reports that ad sets with a diverse mix of images, videos and 9:16 vertical video with audio had a 9.1% lower cost per action than ad sets missing one of those assets, and that ad sets with at least 20% of their assets in 9:16 video with audio saw a 7% lower CPA. These are Meta's figures from its own footnoted studies, not independent benchmarks, and the direction is the point: give the retrieval index different things to choose from.

Consolidation second. Meta's page on managing ad volume says that when an advertiser runs too many ads at once, each ad delivers less often, fewer ads exit the learning phase, and more budget is spent before the system can optimize. Its advice is to combine ad sets, to decrease ads per ad set while keeping diverse assets in each, and to remember that one ad can carry up to 10 creative assets. For reporting by audience, Meta points to breakdowns instead of separate ad sets.

Advantage+ third. For sales, leads and app promotion Meta recommends Advantage+ campaigns, which apply AI across audience, placements, budget and destination. Where you keep a manual structure, Meta recommends A/B testing Advantage+ audience for almost all campaign types except retargeting, and quotes a 7.2% lower cost per result for the Sales and App promotion objectives, 9.7% for Traffic, Engagement and Leads, and 14.8% for Awareness. Advantage+ audience treats age, gender, detailed targeting and included custom audiences as suggestions it can deliver beyond; only locations, minimum age, languages and excluded custom audiences are hard controls.

• An image, a video and a 9:16 video with audio in every ad set
• Fewer ad sets, each with enough budget to exit learning
• Fewer ads per ad set, more varied assets per ad, up to 10
• Breakdowns for reporting, not ad sets
• Advantage+ audience tested against your manual audience, except on retargeting

Since the model chooses among creatives per person, the useful unit of reading is the creative, not the ad set. Meta's Ads Manager shows two statuses for it. Creative limited appears when an ad's cost per result is higher than ads you ran in the past but less than twice as high; Creative fatigue appears when it is at least twice as high. Meta counts all recent exposures of the same image or video, including from other campaigns of your Page, so the same asset in three ad sets fatigues as one.

Meta's recommended responses are to create another ad that is materially different while keeping the original active, to expand the audience, or to try Advantage+ creative. The statuses only exist for ad sets with a single creative, so an account that follows the consolidation advice has to measure fatigue itself: spend, results, CPA, CTR and frequency per asset, over time, across every ad that uses it.

Adrails reads the account that way. Its Creatives page lists each image, video and text asset once, with its spend, results, CPA, ROAS, CTR, frequency and the number of ads that use it across every account, and gives each a signal: winning, fatigued, learning under 1,000 impressions, wasted or steady. Tags in the ad name such as [angle: price] group creatives by angle, hook or persona, and the Creative fatigue automation template posts the ones wearing out to Slack, Discord or email on a schedule.

Do not duplicate ad sets to force spend or to run the same ad against overlapping audiences. Meta's page on auction overlap explains that when two or more of your ads enter the same auction, only the one with the highest total value competes and the others are not considered. Overlap can keep an ad set from spending its budget or from reaching enough results to exit the learning phase, and Meta notes that running the same ads from separate ad accounts may not avoid it, since the delivery system can identify the accounts as one advertiser.

Do not slice audiences into micro segments. Meta's fragmentation page gives the example of three ad sets with the same objective and creative, aimed at people interested in dogs, dog toys and dog collars, and says splitting the audience that way could decrease performance because the total audience is split across several ad sets, each with fewer opportunities to learn. One ad set with all three interests, and breakdowns for reporting, is its recommendation.

Do not keep adding ads either. Too many active ads means each one delivers less often and fewer exit learning, according to the ad volume page. Meta's remedy for a fading structure is to turn off the ad sets that are learning limited or have the fewest results and move their budget to the active one, so that the optimization events land in one place.

• Duplicating an ad set to double spend: the copies meet in the same auctions and only one is considered
• Splitting one interest into several ad sets: each gets fewer events and more time in learning
• Separate ad accounts for the same ads: Meta may still treat them as one advertiser
• Twenty near-identical ads in one ad set: fewer of them exit the learning phase
• Pausing a fatigued original before its replacement delivers: Meta says keeping it active may maximize results

Andromeda is one of several systems Meta has named, and the Meta Ads algorithm in 2026 is the stack they form. In its March 2025 post Meta for Business lists Meta Lattice, a single ranking architecture that replaced many smaller models and that Meta credits with almost 12% higher ad quality and up to 6% more conversions, and sequence learning, which models the order of a person's actions around an ad and added 3% conversions on tested segments.

Above them sits GEM, the Generative Ads Recommendation Model, which Meta's engineering team described in November 2025 as its most advanced ads foundation model, trained across thousands of GPUs on an LLM-inspired paradigm. GEM does not serve ads itself; it teaches the models that do, and Meta reports a 5% increase in ad conversions on Instagram and 3% on Facebook Feed from it. For an advertiser the division of labour is simple: Andromeda chooses the candidates, Lattice ranks them, GEM makes both better over time. Your inputs to all three are the same: the conversion event, the budget and the creatives.

Read the account at the creative level once a week, with a fixed window, and change one thing at a time. Performance is less stable while an ad set is in learning, so the checklist separates reading from acting: most lines end in a note, not an edit.

The table below is the version to start from. Each line names where to look, in Ads Manager or in a tool that reads per creative, and the condition under which to act. The thresholds are Meta's own: twice the earlier cost per result for fatigue, one image, one video and one 9:16 video with audio per ad set, up to 10 assets per ad.

• Creative statuses: any Creative limited or Creative fatigue in the Delivery column
• Spend concentration: the share of spend on the top creative, and whether a new one got any
• Overlap and fragmentation: opportunity score recommendations to combine ad sets
• Formats: an image, a video and a 9:16 video with audio in each ad set
• Learning: ad sets still learning or learning limited after a week