10,000x
Andromeda model capacity
Meta Engineering Blog, Dec 2024
+22%
ROAS with Advantage+ Creative
Official Meta figures
-7%
ROAS post-rollout (n=3,014)
Confect study, March 2026
The context
December 2024: Meta publishes the announcement of Andromeda on its engineering blog, a new ads retrieval system based on custom hardware (NVIDIA Grace Hopper + in-house MTIA accelerators). Model capacity increases by a factor of 10,000 over the old Lattice.
The rollout spreads out: starting in Q1 2025, broad deployment in the summer, global completion in October 2025. That is when PPC agencies start seeing accounts flip overnight. ROAS at 5x dropping to 2x. Performance stable for 3 years collapsing in 6 weeks.
On May 13, 2026, at Meta's Performance Marketing Summit in San Jose, Andromeda is at the center of every product announcement. The recaps from Foxwell Digital, Aspire and Code3 converge on one message: targeting is no longer an input to the system, the creative determines who the ad is served to.
December 2, 2024
Official announcement (Meta Engineering Blog)
Detailed architecture published by 8 Meta engineers. Presented as the biggest delivery system overhaul since Advantage+.
January 2025
Progressive rollout (Q1 2025)
First advertisers flipped. Agencies observe the first shifts in creative vs audience behavior.
October 2025
Full rollout
Meta confirms general availability. Accounts that hadn't yet migrated flip automatically.
May 13, 2026
Performance Marketing Summit (San Jose)
Andromeda becomes the central angle of Meta's product narrative. Foxwell, Aspire, Code3 recaps aligned on the same tactical message.
Source
10,000x increase in model capacity vs the previous retrieval system, +6% recall improvement, +8% ads quality improvement on selected segments, and a 22% increase in ROAS for advertisers adopting Advantage+ creative.
How it actually works
Meta runs four nested AI systems: Andromeda (retrieval), Lattice (ranking), GEM (generative foundation model), and Advantage+ (campaign products). Andromeda is the retrieval stage: out of tens of millions of active ads, it selects a few thousand that might interest a given user.
The key innovation: Andromeda reads your creative at the pixel, video frame and audio cue level, not just metadata. It builds an internal representation of who is likely to enjoy this creative, and that is what it uses to match against users. Your targeting parameters become one input among many, no longer a hard filter.
The inverted model
Si
Before Andromeda
Alors
You pick an audience, you make ads for that audience.
Targeting parameters filter who sees the ad.
Si
With Andromeda
Alors
You feed 8-15 diverse creatives, the system finds the audiences.
Your settings become suggestions; the creative does the targeting.
The community's verdict
Sam Piliero (founder of The Moonlighters, roughly $30M/month under management) wrote one of the most shared LinkedIn posts of the Andromeda era in February 2026. His observation became the reference quote for explaining what changed.
When Meta rolled out Andromeda, I saw hundreds of comments like: "We went from 3x... 6x... even 10x ROAS to barely breaking even." Before: Pick an audience, create ads for that audience. Now: The concept finds the audience. Creative does the targeting. But "creative is the targeting" doesn't mean "upload random ads and pray." It means you need a repeatable concept system. Think in concepts, not winners. Let multiple profitable concepts coexist. Rule of 10k: For every $10K/month in spend → upload 1 new ad concept per week. $100K/month = 10 ads/week.
What the world is saying
Here are the 5 most-engaged LinkedIn posts of the last 6 months on Meta Andromeda. From the contrarian (Jon Loomer saying Andromeda "doesn't want anything") to the purely tactical (Charles Tichenor on testing). Swipe to browse.
What LinkedIn says
Top 5 by engagement · last 6 months
Meta's Andromeda update rewards one thing: creative volume. I know brands quietly producing hundreds of UGC ads without hiring a single creator. $0 spent on talent. The result? They're driving up ROAS faster than their competitors can keep up.
Andromeda doesn't "want" anything. A phrase I keep hearing from advertisers that just hits wrong is "what Andromeda wants." They talk about Andromeda like a living, breathing entity. It's all complete nonsense and a distraction. Don't focus on making Andromeda happy. Focus on PEOPLE.
Meta's Andromeda algorithm changed the game. To win on paid social today, you don't need 5 great creatives. You need hundreds of them, tested, scored, and iterated weekly. Performance creatives are the #1 lever of acquisition, yet teams pour 90% of budget into them with zero system behind it.
Stop Testing So Many Ads on Meta (Andromeda Changed Everything). Most advertisers are still stuck chasing "winning ads," flooding their accounts with endless tests that make things worse. A single "winning ad" can never scale on its own.
Hot take: the biggest competitive advantage on Meta is one thing only. It's your creative volume. Your creative diversity. Your willingness to let things fail fast. Meta's Andromeda engine is a fundamental shift in how advertising works on the platform.
What actually works in 2026
Four tactical changes are validated by just about everyone running a meaningful account (Foxwell Digital, Moonlighters, Wonderful, Marketing Operators):
The 4 adjustments to make
Consolidate your structure
1-3 campaigns per account maximum. Multiple ad sets under a single CBO campaign. The more you fragment, the more you starve the system of signal.
Switch to broad audiences
Wonderful documents +49% ROAS for broad targeting vs lookalike. AI does better than your detailed filters. Lookalikes are still relevant in some verticals (high LTV), but broad is the default now.
Push 8-15 variations per ad set
Not 8-15 near-duplicates: 8-15 genuinely distinct angles, hooks and formats. Andromeda's similarity detection means psychologically redundant variations count as a single creative in the auction.
Measure at the campaign level, not per ad
A creative at 1.5x individual ROAS may be feeding a top funnel that makes another creative at 4x possible. Killing the 1.5x can kill the 4x. The system is holistic now.
8-15
Variations per ad set minimum
Wonderful benchmark, Feb 2026
30-50
Creatives per month (competitive account)
Atria, DTC agencies
+49%
ROAS broad vs lookalike
Wonderful, DTC client data
The typical account before vs after adjustment
| Element | Before Andromeda | After adjustment |
|---|---|---|
| Number of campaigns | 8-15 (fragmented) | 1-3 (consolidated) |
| Main targeting | Detailed audiences + lookalikes | Broad + Advantage+ |
| Creative variations per ad set | 3-5 similar ads | 8-15 distinct concepts |
| Cadence of new creatives | 1-2 per month | 1 per $10k of spend per week |
| Performance measurement | Per individual ad | At campaign/account level |
| Mid-flight changes | Frequent, per ad | Rare, never on winning ads |
The contrarian angle
Meta announces double-digit gains. The ground truth is more nuanced. Confect analyzed 115.7 billion impressions across 3,014 advertisers post-rollout: median ROAS dropped 7%, and top performers (the top decile from before Andromeda) saw their ROAS drop 31%.
PPC Land published an incrementality test showing that Advantage+ generates only 17% of the conversions reported by Meta's attribution system. The gap between reported ROAS and actual incremental ROAS has become a major topic in the community.
Pour
- Meta's AI excels at finding user pockets invisible to manual targeting
- Consolidated structures simplify management (fewer accounts to monitor)
- Diversified creative reduces fatigue and extends concept lifespan
- Under-delivering accounts can unlock additional free reach
Contre
- Loss of control over where and to whom your budget is spent
- Reported ROAS can overstate actual incremental ROAS by 30 to 80%
- Creative production becomes a constant operational load (dedicated team or freelancers)
- High-specificity accounts (niche B2B, hyper-local geo) often lose performance
Source
Median ROAS -7%. Top performers (top decile from before Andromeda) saw their ROAS drop 31%.
What we advise against in agency
This week on your accounts
- Count active campaigns and identify the ones to consolidate
- Audit ad sets with fewer than 5 distinct creatives
- List active detailed audiences and test a broad duplicate in parallel
- Check that you're hitting the minimum creative cadence (Rule of 10k)
- Activate Advantage+ Creative on eligible accounts
- Measure performance at the campaign level for at least 14 days before judging
Verdict
If your Meta Ads account relies on detailed targeting and 3-5 variations per ad set, you're probably already paying the Andromeda tax without knowing it. The transition to a creative-first model takes 2 to 3 months to stabilize, and demands a real operational shift: cadenced creative production, simplified structure, campaign-level measurement.
For most Quebec SMBs under $30,000/month in spend, the most impactful change is to switch to broad audiences and to produce more truly distinct variations. If you don't have the in-house capacity to sustain the creative pace, it's become a strategic decision: either you hire/outsource creative, or you accept a gradual degradation of your performance.
Want a post-Andromeda audit of your Meta Ads account?
Request a free auditFrequently asked questions
Does Andromeda also affect Instagram Ads?
Yes. Andromeda is the retrieval system for the entire Meta family (Facebook, Instagram, Messenger, WhatsApp). Tactical adjustments apply uniformly.
Should we drop lookalikes entirely?
No. Lookalikes remain relevant for high-LTV verticals (B2B, SaaS, premium training) where user quality matters more than volume. But for most e-commerce and SMB service cases, broad beats lookalike by a wide margin in 2026.
How long does Smart Bidding take to readjust after consolidation?
14 days minimum, ideally 21. Any structural change resets part of the learning phase. Plan your consolidations when you can absorb 2-3 weeks of variance.
How do I know if I'm hit by the degradation?
Compare your actual spend over 6 months and look at whether your median ROAS dropped without a major change in settings or product. If yes, and you haven't adapted the creative structure, it's probably Andromeda.
Do you really need 8-15 ads per ad set?
8 minimum, ideally 12-15. And these are distinct concepts, not cosmetic variations. If you produce 15 versions of the same hook with a different color, the system sees them as a single ad.