Guide · Meta ads
Meta Andromeda Explained: The Creative Diversification Playbook
What Meta has actually published about Andromeda and GEM, what's still practitioner guesswork, and a step-by-step way to build a creative library the system can use.
By Chad Westleigh · Updated
What is Meta Andromeda?
Meta Andromeda is the first stage of Meta's ad delivery system: retrieval. Every time someone opens Facebook or Instagram, the system has far more eligible ads than it can score in detail, so Andromeda narrows that pool to a shortlist. Ranking models then predict which shortlisted ad is most likely to drive the result you optimize for, and the auction picks the winner.
Meta's engineering team introduced Andromeda on December 2, 2024 and linked it directly to automation and creative volume. The same post noted that more than a million advertisers had used Meta's generative AI tools to create more than 15 million ads in a month. More ads per advertiser meant retrieval had to get much better at matching ads to people, and Andromeda, which runs on NVIDIA's Grace Hopper Superchip, was built for that.
The practical consequence: the system finds audiences through what is in your ads. Your creative library has become your targeting. For the commercial side (how we plan and score creative before spend), see our Meta ads agency service for DTC brands.
The Meta Andromeda update: what Meta has published
Most Andromeda advice online is inference. These are the three primary sources, with Meta's own figures. Meta has not published independent verification of them.
| Date | Meta source | What Meta said | What it means for advertisers |
|---|---|---|---|
| December 2, 2024 | Engineering at Meta: Andromeda | A personalized ads retrieval engine with a 10,000x increase in model capacity, +6% recall and +8% ads quality on selected segments | The system can weigh far more ads per person, so a large, varied library becomes usable instead of redundant |
| April 22, 2025 | Meta for Business: The Creative Advantage | The focus has shifted from narrow targeting to creative diversification as the best lever to find relevant audiences; Advantage+ creative image generation averaged about 11% higher click-through and 7.6% higher conversion rates | Meta itself now frames creative as the main audience input |
| November 10, 2025 | Engineering at Meta: GEM | The Generative Ads Recommendation Model, a foundation model that passes what it learns to hundreds of ranking models; ad conversions rose 5% on Instagram and 3% on Facebook Feed in Q2 2025 | Ranking learns from creative and conversion data, so signal quality matters as much as the ad |
Meta Andromeda myths vs what's documented
| What you'll hear | What's documented | What we'd do |
|---|---|---|
| Andromeda is a new campaign type you switch on | It's back-end infrastructure; there is no setting | Change the inputs: creative, signal and structure |
| Targeting doesn't matter anymore | Meta calls creative diversification the best lever for finding audiences; location, age and exclusions still apply | Keep hard constraints, drop interest stacking |
| More ads always means more reach | Meta's guidance stresses meaningfully different creative, not volume | Count concepts, not ads |
| There's an official 60% creative similarity cutoff | Practitioner reports cite several thresholds; we haven't found one in Meta's documentation | Treat similarity as a direction: avoid near-duplicates |
Why creative diversification beats variant volume
Creative diversification works because retrieval matches people to ads using signals from the ad itself: who appears in it, what it shows, what it says and how it opens. Ads that share those signals compete for the same people. A genuinely different concept reaches a different pocket of the audience.
The working distinction is between a concept and a variation:
- A concept changes the person the ad is for, the problem or desire it speaks to, the promise or the proof.
- A variation keeps the idea and changes the execution: a new headline, crop, color, first frame or length.
Variations help you improve a concept that works. Only new concepts open new audiences. An account with 40 live ads can have three concepts, and it will behave like an account with three ads.
Concepts that win on Meta usually deserve a place in email and on site too; our jewelry email marketing playbook shows how to carry them into flows.
Meta GEM model: how Andromeda, GEM and Lattice fit together
The Meta GEM model, Meta's Generative Ads Recommendation Model, is a foundation model that improves ads ranking. Meta's engineering team introduced GEM on November 10, 2025 as its largest recommendation foundation model. GEM doesn't score each ad request itself; Meta says running a model that large inside a live request isn't affordable. Instead, GEM passes what it learns, through knowledge distillation, to hundreds of smaller ranking models that serve ads in real time. In an October 13, 2026 update, Meta described GEM as a single central foundation model across its ad surfaces.
Andromeda, GEM and Lattice do different jobs:
- Andromeda is a stage. Retrieval selects a few thousand relevant ads from tens of millions of candidates.
- GEM is a teacher. It trains the ranking models that pick among those candidates.
- Lattice is an architecture. Introduced by Meta AI in May 2023, it replaced many small, siloed ranking models with fewer large ones that learn across surfaces and objectives at once; Meta reported about 8% better ads quality in early Instagram results.
Posts that call Lattice "the auction" or describe Andromeda as GEM's student go beyond what Meta has published.
Andromeda vs GEM vs Lattice at a glance
The first four columns follow Meta's own posts; Meta's reported results for Andromeda and GEM appear in the update table above. The last column is our view.
| System | What Meta says it is | Where it works | First described | What it means for advertisers |
|---|---|---|---|---|
| Andromeda | A personalized ads retrieval engine | Retrieval: picks a few thousand candidates from tens of millions | Engineering at Meta, December 2, 2024 | Distinct concepts give retrieval more ways to match your ads to people |
| GEM | A foundation model that teaches downstream ranking models | Ranking, indirectly, through knowledge transfer | Engineering at Meta, November 10, 2025; updated October 13, 2026 | Clean conversion events shape what ranking learns about every concept |
| Lattice | An architecture that consolidates many ranking models into fewer, larger ones | Ranking, shared across surfaces and objectives | Meta AI blog, May 11, 2023 | Placement-by-placement tactics matter less than strong creative and signal |
A creative diversification playbook in eight steps
This is the sequence we run on Meta accounts in the Andromeda era. It works whether your creative comes from an in-house team, a studio or creators.
Audit live ads by concept
Export every live ad and tag it by person, problem or desire, promise, proof and format. Count distinct concepts, not ads.
Map the gaps
Lay concepts against your customer segments and buying moments. Empty cells become the next brief; crowded cells show where lookalike ads are splitting spend.
Write concept briefs, not ad briefs
One page per concept: who it's for, the single idea, the proof, three hooks and the formats to produce. The template is below.
Produce in batches with real spread
Every batch should add new concepts as well as iterations of proven ones. A batch made only of variations of last month's winner adds volume without adding reach.
Launch into a simple structure
Consolidated campaigns with broad targeting and enough budget per ad set to exit learning, plus a separate testing campaign so new concepts don't disrupt what's working.
Read results by concept
Roll ad-level results up to concept level: spend share, cost per purchase, new-customer share and whether the concept reaches people the others don't.
Refresh before fatigue
Watch frequency and falling click-through by concept, and brief successors before winners fade, timed to your sales calendar.
Retire near-duplicates
Pause variations that split spend with a sibling without adding reach, so retrieval has fewer lookalikes to sort through.
The concept brief template
A concept brief forces the decision that matters under Andromeda: what makes this ad different from everything already live. The example column uses a DTC jewelry brand.
| Field | What to define | Example |
|---|---|---|
| Person | Who the ad is for, in one line | A partner buying a first-anniversary gift |
| Moment | When and why they're buying | Three weeks out, worried about shipping time |
| Single idea | The one thing the ad says | A gift that looks considered and arrives ready to give |
| Proof | Why they should believe it | The gift box on camera, a review describing the reaction |
| Hooks | Three opening frames or lines | The reveal; the box opening; “I almost bought flowers” |
| Formats | Which executions to produce | 15-second video, static, carousel by price point |
| Success signal | How you'll judge it against other concepts | New-customer share and cost per purchase |
Measurement and signal under Andromeda and GEM
GEM, by Meta's description, transfers what it learns from conversion outcomes to the ranking models that score your ads, so missing or double-counted purchase events teach the system the wrong lesson about every concept. Brands in restricted categories face extra limits on what they can send; our CBD marketing agency page covers that case.
Three habits keep the read honest:
- Fix signal first. Run the Conversions API alongside the pixel with deduplication and purchase value before judging any creative test.
- Judge concepts on blended outcomes. Platform ROAS by ad flatters retargeting-style concepts; compare concepts on new-customer share and on blended revenue against total ad spend.
- Test incrementality periodically. A conversion lift study or a geographic holdout tells you whether the program adds sales, not just whether ads collect attribution.
Gift-driven categories need extra care, because demand spikes around dates distort week-over-week comparisons; our DTC jewelry marketing practice plans creative around that calendar.
Common Andromeda-era mistakes to check for
- Dozens of live ads that share one hook, one model and one setting
- Judging a concept on a week of ROAS from one ad set
- Budget split across many small ad sets that never exit learning
- Daily edits that keep resetting learning
- New concepts tested inside the campaign carrying most of the spend
- One winning concept taking nearly all spend, with no successor briefed
- Launching new creative before purchase events are deduplicated
How &LIVY applied creative diversification
Operator Intelligence built an Andromeda GEM ad creative analysis tool for &LIVY, a DTC jewelry brand whose team produces 15–20 new ad creatives every other week. The custom web app scores each new batch against Andromeda and GEM signals before any spend, flags near-duplicates and shows which audiences no ad reaches, so the next brief fills real gaps.
The results: 100% of creative review automated, $428K in annual ad spend efficiency, +22% ROAS and ROI within 2.5 months. The &LIVY case study walks through the build. Brands that want a tool like this for their own team can start with custom AI agent development.
Frequently asked questions
What was the Meta Andromeda update?
Meta announced Andromeda on its engineering blog on December 2, 2024 as a new ads retrieval engine. Advertisers felt the shift through 2025 as Meta pushed Advantage+ automation and, in April 2025, described creative diversification as the best lever for finding audiences.
What is the Meta GEM model?
GEM is Meta's Generative Ads Recommendation Model, a foundation model introduced in November 2025. It doesn't rank ads directly; it transfers what it learns to the smaller ranking models that choose among the candidates Andromeda retrieves.
How many creatives does an account need under Andromeda?
Meta hasn't published a number. What matters is how many distinct concepts are live and whether new ones arrive before the current ones fatigue. A smaller budget can support fewer concepts, so run fewer for longer rather than many that never gather data.
Do we still need separate campaigns for prospecting and retargeting?
In our experience, most accounts now do better with fewer, larger campaigns and broad targeting, plus a separate testing campaign for new concepts. Keep exclusions and hard constraints, and let the creative do the audience work.
Is AI-generated creative enough to diversify?
Usually not on its own. Meta reports gains from Advantage+ creative generation, but most generated assets are variations of an existing ad. New concepts come from customer insight: reviews, support tickets, sales conversations and the moments people buy.
Does this apply to small DTC brands?
Yes, with smaller numbers. A brand spending modestly might run three or four distinct concepts at a time instead of a dozen. The principle is the same: distinct ideas, clean signal and enough budget per ad set to learn. That principle runs through our agentic digital strategy for DTC brands.
Find out how many concepts you're really running
Send us your live ads and your next batch. We'll show you where the library overlaps, which audiences no ad reaches and what to brief next.
Book a scoping callSources
- Engineering at Meta (2024): Andromeda announced December 2, 2024 as Meta's personalized ads retrieval engine; 10,000x model capacity; +6% recall; +8% ads quality on selected segments; runs on NVIDIA Grace Hopper; more than a million advertisers created more than 15 million ads in a month with Meta's generative AI tools
- Meta for Business (2025): Focus has moved from narrow targeting to creative diversification as the best lever to find the most relevant audiences; Advantage+ creative image generation about 11% higher CTR and 7.6% higher conversion rate on average
- Engineering at Meta (2025): GEM (Generative Ads Recommendation Model) transfers knowledge to downstream ranking models; ad conversions +5% on Instagram and +3% on Facebook Feed in Q2 2025
- Jon Loomer Digital (secondary, used only to flag the claim as unverified) (2025): Practitioner reports of creative similarity thresholds (60% and others) are not confirmed in Meta documentation
- Engineering at Meta (2024): Andromeda's retrieval stage selects from tens of millions of ad candidates into a few thousand relevant candidates; retrieval is the first step of a multi-stage ads recommendation system
- Engineering at Meta (2025): GEM introduced November 10, 2025 as the largest foundation model for recommendation systems in the industry; acts through knowledge transfer to hundreds of downstream ranking models
- Engineering at Meta (2026): October 13, 2026 update: GEM moves to a single central foundation model across Meta's ad surfaces; evaluating the large model inside a live ad request is not affordable, so it transfers knowledge to downstream vertical models
- Meta AI (2023): Meta Lattice (May 11, 2023): a model architecture that predicts ad performance across many datasets and optimization goals previously handled by hundreds of separately trained models; ~8% ads quality improvement in early Instagram results
Get the Operator Intelligence briefing
A short, practical note every two weeks on what is working with AI in real businesses. No spam; unsubscribe anytime.
Related
Meta ads agency for DTC brands
Andromeda-era creative strategy and pre-spend creative scoring.
&LIVY case study
How creative analysis delivered $428K in efficiency.
DTC jewelry marketing
Gift-calendar creative and the full jewelry stack.
Jewelry email marketing playbook
Carry winning concepts into Klaviyo flows.