Meta announced Andromeda on December 2, 2024, in a post on its engineering blog. Most advertisers ignored it for months because it read like an infrastructure paper. That was a mistake. By late 2025 the rollout had reached most ad accounts, and the way Meta picks which ads to show people had changed at the root.
I have spent the past year reading everything Meta has published about this system, plus the practitioner writeups and the large-scale spend studies that followed. This article is the plain-language version: what Andromeda is, what it actually changed, what it did not change, and the specific things worth doing differently in your account.
What is Meta Andromeda?
Andromeda is the machine learning system that handles retrieval in Meta’s ads pipeline. Retrieval is the first step in ad delivery. Before any auction happens, Meta has to narrow tens of millions of eligible ads down to a few thousand candidates worth considering for one specific person at one specific moment. Andromeda is the system that does that narrowing.
It is worth being precise here, because most of the confusion around Andromeda comes from skipping this step. Meta’s delivery works in stages:
- Retrieval: tens of millions of ads get filtered to a few thousand candidates. This is Andromeda’s job.
- Ranking: heavier models score those candidates and pick the handful that enter the auction.
- Auction: bid, estimated action rate, and ad quality decide what a person actually sees.
Andromeda did not replace the auction. It replaced the short-listing process that happens before the auction. If your ad never makes the short list, your bid and your budget never get a chance to matter. That is why a retrieval upgrade turned out to be a strategy upgrade.
The numbers Meta published
The engineering post carries specific claims, and they explain why Meta bothered rebuilding this stage. According to Meta’s own figures, Andromeda runs on NVIDIA Grace Hopper superchips and Meta’s own MTIA accelerators, and delivers a 10,000x increase in model capacity at the retrieval stage compared with the prior system. Feature extraction got over 100x faster. Meta reported a 6% recall improvement in retrieval and an 8% improvement in ads quality on the segments it tested.
The context number matters more than any of those: Meta said over one million advertisers were already creating more than 15 million ads a month with its generative AI tools at the time of the announcement. The old retrieval system was built for a world where a big account ran dozens of ads. Andromeda was built for a world where accounts run hundreds, and the system has to find the right one for each person in milliseconds.
The rollout timeline
Dates got muddled in the coverage, so here is the sequence as best as public sources establish it:
- December 2, 2024: Meta publishes the Andromeda engineering post. The system is already live on some traffic.
- Through 2025: phased rollout across accounts and placements. Confect, an ecommerce creative platform, tracked the deployment across its client base in phases it measured at roughly 30%, 70%, and then 100% of traffic.
- October 2025: the rollout reaches substantially all accounts. This is the point most practitioners mark as the “before and after” line, and it is why you saw a wave of “what happened to my CPMs” threads that month.
- Into 2026: Meta layers GEM (its generative ads model) and further Advantage+ automation on top. Andromeda is the foundation these run on.
What actually changed for advertisers
Creative became the targeting input
The old system leaned heavily on your targeting selections to decide who saw your ad. Andromeda reads the ad itself. Its models analyze the visual content, the text, the audio, the concept, and match all of that against person-level interest signals in real time. Meta has said this outright in its advertiser-facing material: the lever has moved from audience selection to creative diversification.
Practically, this means the persona work you used to do in ad set settings now happens inside the creative. An ad that shows a runner, speaks to knee pain, and uses running vocabulary will find runners. You no longer need an interest stack to get it there, and adding one mostly just shrinks the pool Andromeda can work with.
Conceptually similar ads collapse into one slot
This is the change that broke the most playbooks. Andromeda’s models recognize semantic similarity. Twenty ads that are the same concept with different headline colors, crops, or opening words get treated as one idea, and that idea effectively competes for one place on the short list. Practitioners took to calling these clusters “entity IDs.” The term does not appear anywhere in Meta’s documentation, so treat it as shorthand rather than a technical fact, but the observed behavior is consistent: minor variations do not buy you additional retrieval coverage. Distinct concepts do.
The old volume playbook, where you duplicated a winner fifteen times with small tweaks and let the auction sort it out, now works against you. You are not filling the pipeline with fifteen tickets. You are handing the system one ticket with fifteen copies.
Winners emerge faster, and die faster
Faster retrieval cycles mean the system finds your best ad quickly and concentrates delivery on it. It also means that ad exhausts its audience sooner. Fatigue timelines that used to run six to eight weeks now commonly run two to four. Monica Shukla of Mile Marker, writing in AdExchanger in January 2026, described the effect as quicker creative exhaustion and more dynamic optimization, and that matches what most buyers I compare notes with are seeing. I cover the diagnosis side of this in a separate guide to creative fatigue on Meta.
What did not change
Some corrections to the louder claims floating around:
The auction still exists. Bids, budgets, and estimated action rates still decide the final impression. Andromeda decides who gets considered, and that is a big deal, but “the algorithm is entirely new” oversells it.
More creative is still not automatically better. The AdExchanger piece is blunt about this, and the logic holds. If you upload 40 concepts against a budget that can feed conversion data to five, learning fragments and the system struggles to distinguish signal from noise. Creative volume has to scale with budget, or you are just adding noise.
Results did not improve for everyone. The largest independent dataset published on this is Confect’s study of 115.7 billion impressions and 834 million dollars of spend across 3,014 ecommerce advertisers through 2025. Their topline: average ROAS declined 7% across the rollout year, and advertisers who had been top performers before Andromeda lost the most ground, down 31% on average, while bottom performers actually gained slightly. The system compressed the advantage that came from account structure skill. What separates performance now is the creative input, which is exactly the part Meta cannot automate for you yet.
How to tell whether Andromeda changed your account
Plenty of accounts sailed through the rollout without noticing, usually because their creative practice already matched what the new system rewards. If you are diagnosing after the fact, these are the fingerprints I look for when someone shows me an account and asks “was it Andromeda?”
- A performance break around your rollout date, most commonly October 2025, without any corresponding change in settings, tracking, or competition. Fatigue timelines shortening at the same time is the confirming detail.
- Delivery concentrating harder on one or two ads. Where spend used to spread across six or eight ads, it now piles onto the top performer until that ad burns out. This is the retrieval engine finding and exploiting winners faster.
- Duplicated winners underperforming the originals badly. The duplicate-and-scale tactic depends on the system treating copies as new ads. It largely stopped working when semantic clustering arrived.
- Narrow interest-targeted ad sets decaying faster than broad ones. As the creative signal took over, the accounts that kept heavy interest stacks saw their edge evaporate first.
- “Creative limited” and “creative fatigue” flags appearing in Ads Manager at a pace you had not seen in previous years.
Two or three of these together is a reasonably confident diagnosis. The response is the strategy section below, and none of it involves rolling anything back, because there is nothing to roll back to. Meta does not ship a classic mode.
Myths versus what the evidence supports
| Claim you will hear | What the evidence supports |
|---|---|
| “Andromeda replaced the auction” | The auction is intact. Andromeda replaced the short-listing stage that decides which ads reach the auction. |
| “Upload as many ads as possible” | Distinct concepts help. Volume beyond what your budget can feed with conversion data fragments learning and adds noise. |
| “Targeting settings are dead” | Geography, language, and regulatory controls still matter. Interest stacks and lookalikes lost most of their influence to creative-derived signals. |
| “Entity IDs are an official Meta system” | The term is practitioner shorthand for observed clustering behavior. The behavior is real; the terminology is not Meta’s. |
| “Everyone’s performance improved” | Meta reported an 8% ads quality lift on tested segments. The largest independent dataset, Confect’s 115.7 billion impressions, found average ecommerce ROAS declined 7% across 2025, with formerly top-performing accounts losing most. |
| “It’s a temporary test Meta will refine away” | Andromeda is the foundation for GEM and the broader Advantage+ roadmap. This is the direction of travel, and the investment in custom silicon says Meta is not turning around. |
GEM and what comes next
Andromeda is one layer of a larger automation push, and reading it in isolation undersells what Meta is building. GEM, the Generative Ads Model, sits on the creation side: point it at a product URL, give it a prompt, and it assembles campaign creative. Advantage+ automation sits on the buying side, absorbing budget allocation, placements, and bidding. Andromeda in the middle makes the matching between all that machine-generated creative and individual users fast enough to work at scale. Meta’s engineering post projects another 1,000x complexity increase as its MTIA silicon and next-generation GPUs come online, which tells you the retrieval stage will keep getting finer-grained.
The strategic read: every layer of the stack that can be automated is being automated, and the surviving human jobs are the inputs the machine cannot generate. Concretely, that means the ideas inside the creative, the offer, the landing experience, and the measurement discipline to know what is actually working. Social Media Examiner’s 2026 outlook made the same point from a different angle: as GEM automates assembly, post-click experience becomes a bigger share of the competitive advantage. Plan your team’s skills around that split.
What to do differently
Five changes, in the order I would make them:
1. Consolidate your structure. One campaign per conversion objective, broad targeting, Advantage+ settings on. Fragmented ad sets split learning signal and create auction overlap. Structure stopped being the place skill shows up. Keep separations that reflect genuinely different jobs, like prospecting versus retargeting or different markets, and collapse the rest.
2. Rebuild your creative plan around concepts, then variants. Aim for five to ten genuinely distinct concepts live per campaign. Different pain point, different persona, different message structure, different format. Then iterate variants of whichever concepts win. I wrote a full playbook on creative diversity that covers how to tell a concept from a variant, because that distinction now decides your retrieval coverage.
3. Move persona targeting into the ads. Take every audience you used to target with interest stacks and write it into the creative instead. The ad for new parents should look and sound like it was made for new parents. That is the signal Andromeda reads.
4. Feed the system clean data. Conversions API alongside the pixel, full event coverage, deduplication verified. Retrieval quality depends on signal quality. This was true before Andromeda. It is more true now because the system is making finer-grained matching decisions.
5. Plan production capacity before you plan campaigns. Faster fatigue plus a concept-diversity requirement means the binding constraint in most accounts is now how many genuinely different ads you can produce per month. Budget for that the way you budget media.
Frequently asked questions
Is Andromeda the same thing as GEM?
No. Andromeda is the retrieval engine. GEM is Meta’s generative ads model, aimed at creating and assembling ad content. They are complementary layers of the same automation push, but they do different jobs.
When did Meta Andromeda launch?
Meta announced it on December 2, 2024, with the system already live on part of its traffic. The rollout phased through 2025 and reached substantially all accounts around October 2025.
Did Andromeda kill interest targeting?
Interest targeting still exists as a control, but its influence has shrunk. The system increasingly derives audience fit from the creative itself. Most accounts perform at least as well with broad targeting and clearly differentiated creative, with exceptions for regulated categories and geographic restrictions.
Do I need 50 ads per ad set now?
No, and uploading 50 near-duplicates can hurt. What matters is distinct concepts matched to your budget’s ability to feed them data. For most accounts that is five to ten concepts, refreshed as they fatigue.
Why did my performance drop after October 2025?
You are not alone. In the Confect dataset, conversion rates fell 17% on average across the rollout, and previously strong accounts were hit hardest. The common thread among accounts that recovered was a rebuilt creative pipeline: more distinct concepts, faster refresh, and broad targeting. Structure tweaks alone did not bring accounts back.
Does Andromeda affect Google or TikTok ads?
No. Andromeda is Meta infrastructure, covering Facebook, Instagram, and the rest of Meta’s placements. That said, every major platform is moving the same direction, from Google’s Performance Max to TikTok’s Smart+, so the underlying skill shift toward creative strategy travels well.
Is there any way to opt out of Andromeda?
No. It is the delivery infrastructure, not a feature toggle. The controls you retain are your inputs: creative, signal quality, budget, and the remaining audience constraints. Accounts do better treating it as physics to work with rather than a setting to fight.
Sources: Meta Engineering (December 2024 Andromeda announcement), Meta for Business creative diversification guidance, Confect 2025 Andromeda impact study, AdExchanger (January 2026), Jon Loomer Digital, Social Media Examiner.
