Meta's Andromeda Update: What Actually Changed and What to Do About It (2026)
Andromeda rebuilt Meta's ad retrieval engine and turned creative into targeting. What broke in existing accounts, the data behind the ROAS drop, what changed in 2026, and the structure that works now.
if your meta ads performance tanked in late 2025 and hasn't fully recovered, andromeda is probably why. this guide explains what actually changed - not the hype version, the real version - and exactly what to do differently.
what andromeda actually is
andromeda is meta's rebuilt ad retrieval engine. it went live mid-2025 and completed global rollout by october 2025.
here's the thing most guides get wrong: andromeda is not a new targeting system. it's not a button in ads manager. it's the infrastructure layer that runs before any targeting or ranking happens - the system that decides which ads even get considered for a given user out of tens of millions of eligible ads.
the old retrieval system was slow. it couldn't handle the volume of ads being created, especially after meta's generative creative tools started making one ad into hundreds of variations. so meta rebuilt it from scratch. andromeda is reportedly:
- 10,000x more complex than the previous model
- 100x faster at feature extraction
- 3x higher throughput
- +6% improvement in retrieval accuracy
- +8% boost in ad quality
the practical result: meta can now consider far more ads per user, far faster, with far more nuance about what that specific user actually wants to see.
where andromeda sits in the stack
andromeda is one stage of a multi-stage pipeline, not the whole algorithm. meta ads is at least four named systems:
- retrieval (andromeda): picks a few thousand candidate ads out of tens of millions. reads your creative. this guide.
- ranking (lattice, the adaptive ranking model, and hundreds of models taught by gem): scores those candidates on how likely this person is to convert next. reads your conversion signal.
- gem (generative ads recommendation model): the llm-scale foundation model that does not serve ads directly but teaches every ranking and retrieval model what a conversion journey looks like across facebook and instagram.
- auction and pacing: bid x estimated action rate + ad quality. reads your bids, budgets, and learning-phase status.
why it matters: creative fixes only help at the retrieval stage. if your problem is downstream (conversions fell while impressions and ctr held), andromeda is not the layer to fix. the full map, with meta's reported numbers for each system, is in the algorithms behind meta ads.
the fundamental shift: audience-first to creative-first
this is the part that actually matters for your account.
before andromeda: you defined an audience (interests, behaviors, lookalikes). meta found users within that audience and picked the best-performing ad to show them. one or two creatives would "win" and eat most of the budget.
after andromeda: meta reads your creative - copy, visuals, hooks, format, audio - and uses those signals to predict which users across its entire network are most likely to take action. your audience definition is now a starting suggestion, not a hard constraint.
in plain english: your creative is now your targeting.
the copy you write, the hook you use, the problem you describe in the first 3 seconds of video - these signals tell andromeda who to find. if your ad talks about a problem only 35-year-old mothers with back pain have, andromeda finds those people. you don't have to target them manually.
what this broke in existing accounts
1. interest targeting and lookalike audiences
these still work as signals, but they now function as ceilings, not floors. adding narrow interest targeting restricts the audience andromeda can pull from. the advertisers who leaned hardest on detailed targeting saw the worst performance drops.
meta has quietly removed the "narrow" targeting option from more and more accounts. for most accounts, location + minimum age are the only hard constraints left. everything else is a suggestion.
2. the 6-ad-per-ad-set rule
for years meta recommended no more than 6 ads per ad set. that rule is gone. andromeda can handle and actively benefits from more creative variety. 15-20 creatives per ad set is the new recommended range for accounts with sufficient budget.
3. turning off "underperforming" top-of-funnel ads
this is where a lot of accounts got wrecked. andromeda and the ranking layer behind it (gem and the models it teaches, see the algorithms behind meta ads) optimize ad sequences, not individual ads. if you killed a top-of-funnel ad because its in-platform ROAS looked bad, you may have collapsed the sequence below it. brands hit by this typically saw ROAS drop 30-50% before they understood what happened.
4. creative similarity killing delivery
andromeda groups visually similar ads into single entity IDs. if you uploaded 20 ads that all have the same background, same face, same format - meta treats them as one ad with 20 variations. delivery gets severely limited. this is why campaigns with lots of ad variations but no real creative diversity still underperform.
the data: what actually happened
from a study covering 115.7 billion impressions, $834M in ad spend, 3,014 advertisers across 73 countries in 2025:
- overall ecommerce ROAS declined 7% post-andromeda with no sign of recovery yet
- top performers (top third by ROAS in their industry) saw a 31% ROAS collapse - from ~17x to ~11x
- mid and bottom performers barely moved or slightly improved
- andromeda acted as a "great equalizer" - compressing the gap between elite advertisers and everyone else
- CPMs are up 20% year over year (avg ecommerce CPM: $16.80 in 2026)
the brutal part: the advertisers who were winning the most before andromeda got punished the hardest. if your account was highly optimized for the old system, that optimization worked against you.
the upside: advertisers who fully adapted to andromeda best practices report 20-35% higher ROAS vs those still running legacy structures.
what changed in 2026 specifically
february 2026: ad sequencing now available for awareness and engagement campaigns - previously only available for reservation campaigns. you can now control the exact order users see your ads.
early 2026: advantage+ audience became the default campaign creation path. manual audience targeting is now the non-default option you have to opt into.
ongoing: meta consolidated narrow interest categories into broader groups. many granular interests that existed in 2024 are gone or merged.
attribution reporting change: link clicks are now isolated as the only metric in standard click-through columns. broader actions (page likes, comments, short video views) shifted to a 1-day engage-through column. this makes reported ROAS look higher than actual revenue in many accounts - check your column setup.
what actually works now
1. consolidate into fewer campaigns
complicated campaign trees choke andromeda's learning. the algorithm needs concentrated data flow.
the structure that works:
- 1 campaign with campaign-level budget (CBO)
- 1-2 ad sets with broad targeting
- 15-20 unique creatives in the ad set
when data is split across 5-10 ad sets, each gets too few signals to optimize. one broad ad set with 15+ creatives gives andromeda what it needs to find winners fast.
2. treat creative as targeting
before you brief creative, ask: "what specific person does this ad describe?" the more specific the problem your ad articulates, the more specifically andromeda can find the right audience.
vague ads reach vague audiences. specific ads find specific buyers.
this means:
- write copy to one persona, not everyone
- use hooks that call out the exact problem ("tired of ads that worked last year but don't anymore?")
- don't make ads that could be for any product in your category
3. real creative diversity (not just variations)
andromeda's entity ID system groups visually similar ads. changing the headline on the same background image does NOT count as a new creative in andromeda's eyes.
real diversity means:
- different formats (video, static, carousel, reels)
- different hooks (problem/solution, testimonial, demo, trend)
- different visual styles (UGC, studio, animation)
- different messengers (different people, different settings)
the practical test: if you can't clearly articulate what's different about two ads beyond the copy, andromeda sees them as the same ad.
4. nail the first 3 seconds
andromeda scores video ads' first 3 seconds separately. a strong hook doesn't just stop the scroll - it signals to andromeda what type of user this ad is for. weak hooks limit both CTR and delivery.
5. wait longer before making changes
meta now recommends 7+ days before making changes (up from 3-4 days). andromeda and gem need longer learning periods. cutting campaigns early is one of the most common mistakes killing performance right now.
6. clean conversion data is critical
andromeda's retrieval is only as good as the conversion signal you're sending it. CAPI setup, deduplication, and match quality matter more than they ever did. the system needs to know exactly who converted to find more people like them.
if your event match quality (EMQ) score is below 6, fix that before anything else.
7. keep winning ads running longer
ad fatigue under andromeda is real - a single concept burns through its audience in 2-3 weeks now vs 6+ weeks pre-andromeda. but don't turn off ads that are still delivering. add new creative alongside winners rather than replacing them.
treat your creative library like a portfolio: every 1-2 weeks, cut underperformers and add fresh concepts. don't touch what's working.
what to stop doing immediately
- stop layering interest stacks on top of each other
- stop duplicating ad sets to "test" audiences
- stop turning off ads the day they stop scaling
- stop running fewer than 10 creatives in an active ad set
- stop making 20 "variations" that all look the same
- stop checking performance before the 7-day learning period ends
- stop optimizing for clicks if you want conversions
the orbit agent instructions
{
"guide": "meta-andromeda-2026",
"agent_tasks": {
"audit_campaign_structure": {
"check": "count active campaigns, ad sets per campaign, ads per ad set",
"flag": "more than 3 campaigns per objective, more than 2 ad sets per campaign, fewer than 10 ads per ad set",
"recommend": "consolidate to 1 campaign, 1-2 broad ad sets, 15-20 creatives"
},
"audit_targeting": {
"check": "identify ad sets using interest targeting or lookalike audiences as primary audience",
"flag": "any ad set with narrow interests enabled, any lookalike audience under 1M",
"recommend": "switch to advantage+ audience or broad (location + age minimum only)"
},
"audit_creative_diversity": {
"check": "categorize active creatives by format, hook type, visual style, messenger",
"flag": "more than 50% of creatives sharing same format or visual style",
"recommend": "add creatives across at least 3 formats and 3 hook types"
},
"audit_conversion_signal": {
"check": "pull event match quality score from events manager",
"flag": "EMQ score below 6.0",
"recommend": "audit CAPI setup, check deduplication, verify pixel firing correctly"
},
"monitor_creative_fatigue": {
"check": "weekly frequency, CTR trend, CPM trend per creative",
"flag": "frequency above 3.5, CTR declining more than 20% week over week",
"recommend": "add 3-5 new creatives with different hooks/formats, do not turn off existing winners"
}
}
}
the bottom line
andromeda is not going away and meta is continuing to tighten it throughout 2026. the gap between accounts that adapted and accounts still running legacy structures is widening every quarter.
your job changed. it's no longer about finding the right audience. it's about creating the right signal for meta to find the audience for you.
creative is the new targeting. the accounts winning in 2026 have accepted that and built systems around it.
built by orbit - the performance marketing intelligence layer. questions? jose@orbitllm.com