Everything you need to know about Andromeda, Meta's advertising AI

If you’ve been running Meta campaigns for a while, you’ve probably noticed the rules have changed. Detailed audiences matter less. Manual segmentation underperforms. Creative, on the other hand, moves everything. According to Meta’s engineering blog, Andromeda’s neural network is 10,000 times more complex than its predecessor and runs feature extraction 100 times faster — and it uses your creative content as the primary signal for finding the right audience (Meta Engineering Blog, December 2024).

There’s a technical reason behind that shift. It’s called Andromeda.

To understand how delivery decisions get made downstream, it helps to know how Meta’s attribution system interacts with Andromeda-driven delivery.

---

Key Takeaways

• Andromeda is Meta’s ad retrieval engine, announced December 2024, that filters millions of ad candidates down to thousands before the auction. Your ad must pass it to compete, regardless of your bid (Meta Engineering Blog, 2024).
• The system’s neural network is 10,000x more complex than its predecessor, with feature extraction latency 100x faster. Meta reports a +22% ROAS improvement for Advantage+ Creative users — treat as directional (Meta for Business, 2025).
• Creative is now the targeting mechanism. Andromeda reads visuals and copy to find the right audience, making diverse creative concepts more important than narrow audience segmentation.
• Hyper-segmented accounts with many narrow ad sets perform worse under Andromeda. Consolidated campaigns with broad targeting and diverse creative outperform them consistently.

Contents

1. What is Andromeda?
2. The problem it was built to solve
3. How it works
4. Andromeda inside Meta’s full AI stack
5. The numbers Meta has published
6. What this changes for advertisers
7. The criticisms worth taking seriously
8. The essentials for now

---

What is Andromeda?

Andromeda is Meta’s ad retrieval engine — the first filter that decides which ads from tens of millions of eligible candidates even get a seat at the table before the auction starts. According to Meta’s engineering blog (December 2024), Andromeda was built to solve a specific problem: Meta’s old retrieval system could evaluate only a few hundred ad candidates per request, meaning the vast majority of potentially relevant ads never competed at all.

Andromeda is not a bidding system. Not a ranking system. It’s the gatekeeper.

Here’s the thing most people miss: your ad doesn’t compete with every other ad on Meta. It first has to pass Andromeda’s filter. If it doesn’t, it never shows up, no matter how aggressive your bid. The auction happens downstream. Andromeda decides who gets to play.

Meta announced it on December 2, 2024, on their engineering blog. Global rollout finished October 2025.

---

The problem it was built to solve

Meta’s ad volume created a retrieval problem at a scale the old system simply couldn’t handle. According to Meta for Business, over a million advertisers are now using Meta’s generative AI tools to produce more than 15 million ads per month — a volume that would overwhelm any retrieval architecture not designed for it from the ground up.

Beyond volume, the old system had a personalisation ceiling. It ran on isolated models — separate ones for video, conversions, reach — that didn’t talk to each other, plus a pile of hand-coded rules that capped how sophisticated the targeting could actually get. The CPU components couldn’t handle real-time feature extraction at scale. It was slow, fragmented, and increasingly out of its depth.

Andromeda is the answer to all of that. One system. End-to-end. Built from the ground up.

The scale problem here is worth pausing on. With 15 million ads per month and hundreds of millions of daily users, the old retrieval system was essentially making random selection calls on most of the available inventory. Andromeda doesn’t just improve relevance — it fundamentally changes which ads have any chance of being served. Your ad being correctly configured isn’t enough anymore. It needs to be indexable and retrievable before it can even reach the auction.

---

How it works

The core mechanic is embeddings — a technique that maps users, contexts, and ads into the same high-dimensional mathematical space. According to Meta’s engineering blog (December 2024), relevance in Andromeda is literally measured as geometric distance: ads whose coordinates sit closest to a user’s coordinates are the ones that get surfaced.

What makes this fast enough to run at Meta’s scale is the hierarchical index. Rather than scanning every ad in sequence, the index prunes the search space non-linearly. The clever part: the index and the retrieval model are trained together, so the structure of the index reflects what the neural network actually considers relevant — not a separate step bolted on after the fact.

The result is sublinear inference cost. Retrieval time grows slower than the total number of ads. That’s the only way you can go from evaluating hundreds of candidates per request to thousands without the whole thing grinding to a halt.

The neural network itself is 10,000 times more complex than the prior system. That’s not a marketing number — it’s the architectural headroom that Andromeda’s sublinear approach unlocked. And it runs on NVIDIA Grace Hopper Superchips, which store all precomputed ad embeddings directly in GPU memory, cutting feature extraction latency by 100x compared to the CPU-based setup it replaced. End-to-end inference throughput improved 3x.

---

Andromeda inside Meta’s full AI stack

Andromeda doesn’t operate alone. It’s one layer in a four-part system that Meta has been assembling since 2023, each component addressing a different part of the ad delivery problem (Meta Engineering Blog, 2024-2026).

GEM (Generative Ads Recommendation Model, launched November 2025) is the brain. It trains on both ad data and organic engagement from across all Meta surfaces, then uses knowledge distillation to teach the downstream systems what genuine relevance looks like — not just what gets clicked. GEM doesn’t serve a single ad. It makes the other models smarter.

Lattice (May 2023) was the first major piece. It replaced hundreds of siloed, objective-specific models with a single unified ranking architecture that learns across surfaces. What works on Facebook informs Instagram. That was a big deal when it shipped.

Andromeda sits before all of that. Whatever Lattice ranks, Andromeda selected first.

UTIS (User True Interest Survey, January 2026) is the calibration layer — direct satisfaction surveys on a 1-5 scale, rather than inferring satisfaction from engagement. It pushed prediction precision from 48.3% to 63.2%, which matters because engagement was always a noisy proxy for whether someone actually wanted to see an ad.

The flow looks like this:

GEM (teaches via knowledge distillation)
    ↓
Lattice (coordinates and ranks) ← UTIS (calibrates with real satisfaction data)
    ↓
Andromeda (retrieval: millions → thousands)
    ↓
Ranking and auction (thousands → one)

Zuckerberg’s stated goal is full automation — advertisers supply a budget and a business objective, the system handles everything else. This stack is the infrastructure for that bet.

---

The numbers Meta has published

Take these with appropriate scepticism — they’re Meta’s own figures, from their own press releases. That said, the engineering metrics (latency, throughput, model capacity) are harder to inflate than conversion metrics, per the analysis in Search Engine Land (2025).

Metric	Result
Model capacity vs. prior system	10,000x
Feature extraction latency vs. CPU	100x faster
End-to-end inference throughput	3x improvement
Inference efficiency (model elasticity)	10x
Retrieval recall improvement	+6%
Ad quality improvement (selected segments)	+8%
ROAS improvement for Advantage+ Creative users	+22%
Conversion lift from image generation	+7%
Ad quality improvement (Lattice)	+12%
Conversion improvement (Lattice)	+6%
GEM efficiency vs. original ranking models	4x
Instagram conversions (GEM)	+5%
Facebook conversions (GEM)	+3%
Simplified campaign structure	+17% conversions, -16% CPA

The engineering numbers are plausible and internally consistent. The ROAS and conversion figures — especially the +22% for Advantage+ Creative — should be treated as directional rather than precise, for reasons I’ll get into below.

---

What this changes for advertisers

The most significant practical shift: creative is now the targeting mechanism.

Andromeda uses creative content as the primary signal for finding the right audience — not the demographic boxes you tick. The system reads your ad’s visuals, copy, hooks, and format to infer who it should reach. Narrow audience constraints don’t make it smarter. They limit the data it can learn from.

I’ve been running accounts under Andromeda since the rollout completed in October 2025. The shift that’s hardest to accept for experienced Meta advertisers is this: the careful audience segmentation work that used to move the needle — layered interests, tight demographic brackets, custom exclusions — now mostly gets in the way. The accounts that improved the most after Andromeda were the ones where we stripped back the structure and let creative do the work. One client went from 12 ad sets across 4 campaigns to 2 ad sets in 1 campaign. ROAS improved 30% in the first month, not because the creative was better, but because the algorithm finally had enough data to learn from.

That changes what good campaign management looks like. Ten variations of the same angle with different headlines isn’t creative diversity — it’s the same ad with a hat on. The system needs genuinely different creative concepts to find different audience pockets. Different emotional hooks. Different framings of the problem. Different formats. Practically, that means running 10 to 20 distinct creative directions per campaign if you’re serious about it.

The other structural shift is consolidation. Hyper-segmented accounts — lots of narrow campaigns, tightly defined audiences, fractured budgets — perform worse under Andromeda than they did under the old system. The model learns faster when it has more data in one place. One or two campaigns, broad targeting, diverse creative in a single ad set. That’s the playbook now.

One detail that gets overlooked: different creators are treated as distinct entity IDs in Andromeda’s index. Running the same creative concept through a different creator unlocks different audience segments the original ad couldn’t reach. That makes creator partnerships a targeting strategy, not just a brand awareness move.

And don’t touch campaigns constantly. Every significant change resets the learning phase. Meta’s own guidance is a minimum of one week — or 50 to 75 conversions — before drawing conclusions or making adjustments. Daily optimisation is mostly noise at this point.

For a closer look at the creative side of this problem, see how ad fatigue interacts with Andromeda’s learning cycle.

---

The criticisms worth taking seriously

Meta’s communication around Andromeda is, predictably, bullish. Here’s what the more sceptical reading looks like.

Attribution is broken in ways that flatter Meta. An incrementality study found Advantage+ was responsible for only 17% of the conversions Meta’s attribution system claimed credit for. The gap exists because Meta counts likes, shares, and saves as conversion-adjacent signals within attribution windows — so an ad that someone passively scrolled past can get credited for a purchase made days later. The system looks better on paper than it is in reality.

A whistleblower alleged Meta inflated ROAS figures for Shops ads by including shipping costs and taxes as revenue. Internal reviews reportedly found 17-19% inflation in early 2024. Meta disputed this. Make of that what you will.

You’ve lost most of your diagnostic tools. When performance drops, you can’t tell if it’s creative fatigue, audience saturation, a platform-side algorithm change, or something else. Meta removed detailed targeting exclusions in January 2025. Dynamic Media was switched on by default in September 2025. The system now automatically spends up to 5% of your budget on placements you explicitly excluded. The black box gets blacker every quarter.

The system can be volatile. CPMs spiked 10x across multiple accounts in February 2024. In April 2024, a system incident burned through full daily budgets in a matter of hours across thousands of advertisers. These aren’t edge cases — they’re documented incidents that Meta largely brushed off.

There’s a structural conflict of interest that nobody’s fully solved. Meta’s AI is simultaneously supposed to optimise for your results and maximise Meta’s ad revenue. Those goals align most of the time, but not always — and you have no independent way to verify which one is winning on any given day.

None of this means Meta Ads don’t work. For most ecommerce and lead gen setups, they still do. But “trust the algorithm” is better advice when you have independent measurement to sanity-check what the algorithm is telling you. Run incrementality tests. Track view-through and click-through attribution separately. Don’t let Meta grade its own homework unchallenged.

If the volatility described above has affected your account, the practical next step is understanding how to reduce Meta Ads costs while maintaining performance.

---

The essentials for now

Andromeda is the most significant change to Meta’s advertising infrastructure since the auction system launched. Not an incremental improvement — a complete architectural rewrite of the first filter in the ad delivery process.

For advertisers, the message is simple: audience strategy is being replaced by creative strategy. The system that decides which ads compete reads your creatives, not your segmentation settings.

That doesn’t mean surrendering to the system uncritically. The criticisms around attribution, opacity, and volatility are real. But ignoring Andromeda — running hyper-segmented campaigns with repetitive creatives — means operating on an outdated understanding of the system managing your budget.

If you want to talk through how to adapt your campaign structure to this, book a free consultation.

---

Frequently Asked Questions

Do I need to change my campaign structure because of Andromeda?

Almost certainly yes, if you’re still running hyper-segmented campaigns with many narrow ad sets. According to Meta’s own data, simplified campaign structures produce +17% more conversions and -16% lower CPA compared to fragmented account structures. The practical direction is fewer campaigns, broader targeting, and more diverse creative concepts within each ad set — letting Andromeda’s retrieval system do the audience-finding work.

How many creative variations should I be running per campaign?

The guidance that’s emerged since Andromeda’s rollout is 10-20 distinct creative directions per campaign — not variations of the same concept. According to the analysis in Search Engine Land (2025), Andromeda reads creative content to identify audience pockets. Ten versions of the same angle with different headlines gives the system no new information. Different emotional hooks, formats, and framings unlock genuinely different audience segments.

Is the +22% ROAS improvement Meta claims for Advantage+ Creative reliable?

Treat it as directional, not precise. The engineering metrics Meta published (latency, throughput, model capacity) are independently verifiable and internally consistent. The conversion and ROAS figures — including the +22% for Advantage+ Creative — come from Meta’s own measurement systems, which have documented attribution inflation issues. An incrementality study found Advantage+ was responsible for only 17% of the conversions Meta’s attribution claimed. Use independent measurement tools to verify claimed improvements.

How does Andromeda interact with the Meta learning phase?

Andromeda makes the learning phase more important, not less. The system needs to build ad embeddings based on your creative and accumulate user-response signals before it can efficiently retrieve your ad for the right users. Meta’s guidance of waiting 7 days or 50-75 conversions before making changes still applies — but under Andromeda, significant structural changes (audience changes, new creative concepts) also reset the embedding calibration, not just the bid learning. Per Meta for Business, minimising edits during the learning phase is more consequential than it was under the old system.

Should I still use interest-based targeting with Andromeda?

Broad targeting outperforms narrow interest stacking for most advertisers under Andromeda. The system uses creative content — not your audience settings — as the primary signal for finding the right users. Narrow interest constraints reduce the data pool the algorithm learns from without meaningfully improving targeting precision. The exception: retargeting campaigns, where audience definition still matters because you’re targeting specific people (site visitors, past purchasers) rather than letting the algorithm prospect.

---

Updated: March 2026. Sources: Meta Engineering Blog (December 2024), Meta for Business, Search Engine Land, Triple Whale.

---

Sources

1. Andromeda: Meta’s new ads retrieval system — Meta Engineering Blog
2. Meta for Business — Advantage+ Creative and campaign structure data
3. Meta AI advertising stack updates — Search Engine Land
4. Andromeda and Meta’s ad AI stack explained — Triple Whale
5. Meta’s GEM model announcement — Meta Engineering Blog
6. UTIS: User True Interest Survey — Meta Engineering Blog
7. Lattice unified ranking architecture — Meta Engineering Blog