The Brief Is Your New Targeting Strategy: What Meta’s Andromeda Actually Changed

How Meta Andromeda’s retrieval layer moved the targeting decision from the ad set to the asset — and what marketing leaders need to rethink now.

TL;DR

• Andromeda is Meta’s ad retrieval engine, launched December 2024. It narrows tens of millions of candidate ads to a few thousand before ranking models select what actually serves.
• Andromeda is not the whole story. Retrieval is stage one. Ranking was rebuilt separately with GEM, Meta’s Generative Ads Recommendation Model, detailed in November 2025. Most industry coverage conflates the two.
• The practical consequence: the targeting decision migrated from the ad set to the creative asset. What an ad looks like, says, and evokes is now a primary input into who sees it.
• The strategic consequence: the audience decision moved out of the media plan and into the brief. Whoever writes the brief is now doing the targeting.
• For regulated verticals, the constraint differs by category. Financial services lost the ability to tell Meta who to find. Healthcare lost the ability to tell Meta what a good outcome looks like. The first is a creative problem. The second is a measurement challenge that the industry faces, and creative won’t fix it.

Meta Andromeda is a personalized ads retrieval engine, announced by Meta’s engineering team on December 2, 2024. Retrieval is the first stage of Meta's multi-stage ads recommendation system: it selects from tens of millions of ad candidates down to a few thousand relevant candidates, which larger ranking models then evaluate to determine what a person actually sees.

For marketing leaders, the important shift isn’t just how Meta’s system works. It’s what that system changes about the relationship between audience strategy, creative development and media activation.

Here’s the distinction most coverage gets wrong: Andromeda handles retrieval. It does not handle ranking.

Ranking was rebuilt separately. In November 2025, Meta detailed GEM (Generative Ads Recommendation Model), described as the largest foundation model built for recommendation systems, trained at LLM scale across thousands of GPUs. Meta reported GEM driving a +5% increase in ad conversions on Instagram and +3% on Facebook Feed, and characterized it as roughly 4x more efficient at converting data into performance than its predecessor ranking models. Meta published a follow-up in August 2026 confirming this isn’t finished. Over twelve months it quadrupled the compute behind GEM and doubled how efficiently that compute is used — meaning the system is improving at a faster rate than it was a year ago, and will keep changing under your campaigns.

Put simply, Andromeda decides which of your ads get a ticket to the auction. GEM decides which one wins it.

Both were rebuilt inside eighteen months. That compounding is why campaigns that worked in 2024 stopped working in 2025 without any change on the advertiser’s side.

What actually changed for media buyers before and after Meta implemented its ad retrieval engine?

Meta’s retrieval system maps users, contexts, and ads into dense representations and matches on semantic proximity. That means the machine-readable content of an ad — the imagery, the setting, the casting, the pacing, the color palette, the on-screen text, the offer, the tone — functions as the targeting instruction.

Put plainly: a media buyer used to describe an audience in a targeting interface. Now the creative describes the audience by depicting it.

An ad set in a suburban kitchen with a woman in her fifties reaches a different pool than the same offer shot in an apartment with a couple in their late twenties. Not because anyone selected an age range. Because the model inferred one.

This shift is real. But two caveats keep it honest:

First, media didn’t lose the job — the job moved. Signal architecture (Conversions API implementation, value-based bidding, event hygiene), account consolidation, exclusion logic, budget strategy, incrementality measurement, and creative governance are all still media disciplines. They simply sit upstream and downstream of the ad set rather than inside it. Anyone telling clients that media buying is now “just creative” is going to be wrong in an expensive way when the signal breaks.

Second, creative is a powerful input, not the only input. User history, engagement sequence, social proof, landing page experience, and conversion signal all feed the same system. Creative happens to be the lever advertisers control most directly.

Here is the part that deserves more attention than it gets.

If creative encodes the audience, then the quality of the audience thinking upstream of the creative determines the quality of the targeting. Not the polish of the deck. The specificity of the understanding. The strategist’s brief is now the highest-leverage document on a given account.

For marketing leaders, that changes where the targeting decision gets made. It doesn’t start when a media buyer selects an audience. It starts when the organization defines who it is trying to reach — and gives creative enough specificity to make that audience recognizable to the platform.

The brief now has to answer questions it may not have carried before.

• Who buys this, and what does their life physically look like?
• What do they wear, drive, cook, watch, and scroll past?
• What environments do they occupy, and what would feel staged to them?
• What does the category get wrong about them that makes them tune out?
• What do they aspire to that they wouldn’t say out loud?
• Where do they live, and what does that place look like on camera?

Every answer becomes a visual, casting, location, or copy decision — and every one of those decisions becomes a retrieval signal. Vague briefs now produce vaguely targeted media. That was always true in spirit. It is now true in mechanism.

Want to put that thinking into practice? Download Lewis’ Andromeda-Ready Brief, an actionable tool for teams looking to build greater audience specificity into the briefs that guide their creative work.

The practical reframe for agency teams: the brief is the targeting spec. Treat it with the rigor formerly reserved for the media plan.

This is where most Andromeda content stops being useful, because it’s written for e-commerce. Regulated categories face structurally different constraints — and they aren’t the same constraint.

Financial services faces the tightest formal restrictions. Meta expanded the former Credit special ad category into Financial Products and Services, with enforcement beginning January 2025. Advertisers who must declare the category face:

• Age locked to 18–65+, with no narrowing
• All genders, no exclusions
• No income, financial-behavior, or detailed-interest targeting
• No location exclusions, with a minimum targeting radius
• No lookalike audiences — Special Ad Audiences, the earlier substitute, were discontinued in 2022

Financial services advertisers lost the audience controls first, which means they’ve been forced into creative-led targeting longer than anyone else. The winners in this vertical have already built the muscle.

Healthcare is not a special ad category, and conflating the two leads to the wrong playbook.

The constraint is data sharing…Nevermind HIPAA Bulletin risk mitigation, Meta’s enforcing its own governance.

Beginning January 2025, Meta rolled out restrictions on categories of websites and apps assigning data sources to restriction tiers. Broad multi-service health systems typically land in a middle tier with limits on mid- and lower-funnel optimization. Condition-specific properties can lose mid- and lower-funnel event tracking entirely, including custom conversions.

The implication is severe and specific: healthcare brands are being asked to feed a hungrier AI at exactly the moment they’re allowed to feed it less. Creative diversity still helps retrieval. But if the model cannot see appointment requests or form fills, optimization drifts toward proxy events that don’t correlate with revenue.

For healthcare organizations, the response is not primarily a creative response. It’s a measurement response: audit your data source categorization and restriction tier in Events Manager before anything else, appeal misclassification where warranted, rebuild the optimization target around permitted upper-funnel events, and reconstruct the lower funnel through CRM-side matching, media mix modeling, and geo-based incrementality testing rather than platform-reported conversions.

Neither is a designated special ad category in the general case — though student lending pulls education advertisers into Financial Products and Services, and any employment or housing adjacency triggers the corresponding category.

That means most telecom and education advertisers retain full signal and full targeting flexibility, competing in an environment where regulated peers cannot.

The opportunity is asymmetric and largely unexploited. For telecom, the creative lever is regional and life-stage specificity — moving households, students, new families, small business owners — expressed visually rather than through geo-fencing.

For Education, it's program-level and life-stage-level creative variety, where a single institution can legitimately produce dozens of genuinely distinct concepts because it genuinely serves dozens of distinct people.

1. Rewrite the brief template. Add a section that forces visual specificity about the audience: environment, casting, wardrobe, objects, rejected clichés. If a creative team can’t shoot from your brief, the model can’t target from your ad. Download our 6-question tool for building audience specificity into your brief.
2. Change what “creative volume” means. Stop counting assets. Count concepts. Distinct persona, distinct desire, distinct awareness stage, distinct format. Five genuinely different ideas will outperform fifty resizes of one.
3. Move measurement upstream before it becomes a crisis. Especially in healthcare. Platform-reported conversions are a weakening currency in regulated categories; incrementality and MMM are the replacement, and they take months to stand up.
4. Consolidate account structure. Fragmented ad sets starve the model of the learning volume it now depends on. Fewer, broader pools with more distinct creative is the current shape of a well-built account.
5. Restructure the pod. Media strategy, creative strategy, and brand strategy can no longer work in sequence with handoffs. The targeting decision is now made in the room where the concept is chosen, which means media has to be in that room.
6. Re-brief your teams on what “targeting” means. Many marketers still evaluate media partners on audience sophistication. That conversation needs to be reset.

What is Meta Andromeda?
Andromeda is Meta’s personalized ads retrieval engine, launched December 2024. It selects a few thousand relevant ad candidates from tens of millions before ranking models determine which ads serve. Meta reported a 10,000x increase in retrieval model capacity, a 6% recall improvement, and an 8% ads quality improvement on selected segments.

Is Andromeda the same as GEM?
No. Andromeda handles retrieval — deciding which ads are eligible. GEM, Meta’s Generative Ads Recommendation Model detailed in November 2025, handles ranking — deciding which eligible ads win. They are separate systems that were rebuilt in parallel.

Did Andromeda eliminate audience targeting on Meta?
No. Targeting inputs still exist and still function, including geography, custom audiences, exclusions, and detailed targeting where permitted. What changed is their relative weight. Creative attributes now carry substantial influence over who is retrieved, which makes the creative asset a primary targeting mechanism rather than an execution of one.

Does creative volume still work on Meta?
Volume of near-identical variants does not work the way it did. Andromeda’s hierarchical index groups semantically similar ads, so minor variations on one concept are unlikely to be evaluated as independent opportunities. Diversity of concept, not quantity of asset, is the current lever.

How does Andromeda affect healthcare advertisers?
Indirectly, and less than data restrictions do. Healthcare is not a special ad category, but Meta’s data source categorization, rolled out from January 2025, restricts mid- and lower-funnel event sharing for many health properties. The binding constraint for healthcare is signal availability, not targeting control.

Which industries are Meta special ad categories?
Housing, employment, financial products and services (expanded from credit, enforced from January 2025), and social issues, elections and politics. Advertisers in these categories face restrictions on age, gender, location, detailed targeting, and lookalike audiences.

None of this makes anyone less necessary. It makes the handoffs matter more. Media owns the plumbing and the proof. Creative owns the targeting, like it or not. Strategy owns the thinking both of them are working from.

For marketing leaders, the question isn’t whether media, creative or strategy matters more. It’s whether those disciplines are working from the same audience understanding and sharing responsibility for the outcome.

The agencies struggling right now aren't the ones with bad media people. They’re the ones where those three skillsets still work in sequence instead of in the same room.