The answer in brief
- Andromeda 2.0 is a nickname. Meta calls it the next generation of Meta Andromeda and described it in an engineering post on 28 October 2026.
- Andromeda is the retrieval stage of Meta’s ad system. It shortlists ads for each impression on Facebook and Instagram before the ranking models choose between them.
- The new version scores every eligible ad. The 2024 version searched an index selectively. Meta reports more than one billion ads scored per second on a single GPU server, which it puts at 100 times the 2024 system.
- Meta credits Andromeda with 2.4% more conversions on Facebook since 2025. Meta gives no baseline or test method for that figure. Separately, it reports 0.24% more conversions on Instagram Reels from Q3 2026 Andromeda improvements.
- There is nothing to switch on. The system already runs in production, and your targeting rules still apply before any ad is scored.
- Our advice: leave your account structure alone, put the effort into ads that differ in idea and not only in trim, and judge the result on your own cost per customer.
What is Andromeda 2.0?
Andromeda 2.0 is the advertising industry’s name for the rebuilt version of Andromeda, the system Meta uses to shortlist ads for each person on Facebook and Instagram. Meta does not use the name. Its engineering post of 28 October 2026, which credits twelve authors, calls the system the next generation, or the 2026 version, of Meta Andromeda. Advertisers were already using the “2.0” nickname before Meta published its October post.
Andromeda does one job, called retrieval. Each time an ad slot opens, Meta has far more eligible ads than it can examine in detail. Retrieval cuts that pool to a shortlist in milliseconds. Larger and slower ranking models then score the shortlist closely, and the auction decides which ad is shown.
The original Andromeda post of 2 December 2024 put numbers on that cut: tens of millions of candidate ads reduced to a few thousand. That is why retrieval matters to an advertiser. An ad that misses the combined retrieval shortlist cannot reach ranking, whatever its bid.
Andromeda is also not Advantage+. Advantage+ is the set of automated campaign tools you see in Ads Manager. Andromeda sits underneath every campaign, automated or manual, and has no control in the interface.
What did Meta change in October 2026?
Meta rebuilt both the model that judges ads at retrieval and the machinery that runs it, so that every eligible ad now receives a score. The 2024 system arranged ads in a layered index and spent its computing only on the most promising branches. The 2026 system has no need to skip any.
The model changed in three places:
- Several models of you, not one. The old design summarised each person with a single model. The new one uses a group of specialised models, which Meta calls experts, so that different sides of a person’s interests are held separately. They use a person’s history of interacting with ads, patterns among groups of users, signals about meaning derived from language models, and engagement across Meta’s surfaces.
- Ads read in context. Each ad is represented by its content combined with the circumstances of the moment it might be served.
- A router between the two. For each pairing of person and ad, a gating network calls on only the most relevant experts. This is the mixture-of-experts design used in large language models, applied to ad retrieval.
The machinery changed too. Checking an advertiser’s targeting rules, scoring the ads and picking the top candidates used to be split across separate machines, and at this scale the network between them became the bottleneck. All three steps now run on one GPU server, which Meta says cut network traffic about 100 times.
If the table extends beyond the screen, scroll sideways to view all columns.
| Feature | Andromeda, December 2024 | Andromeda, October 2026 |
|---|---|---|
| How it searches | Andromeda, December 2024A layered index, computing only on the most relevant branches | Andromeda, October 2026Scores every eligible ad in the pool |
| Model of the person | Andromeda, December 2024One user model | Andromeda, October 2026A group of specialised experts with a router |
| Throughput Meta reports | Andromeda, December 2024Tens of millions of ads per second per GPU card | Andromeda, October 2026More than one billion ads per second on a single GPU server |
| Computing per prediction | Andromeda, December 2024The baseline | Andromeda, October 2026About 300 times fewer operations |
| Targeting checks | Andromeda, December 2024On separate machines from scoring | Andromeda, October 2026On the same GPU as scoring |
| Result Meta cites | Andromeda, December 20246% better retrieval recall across the Facebook and Instagram deployment, with 8% better ad quality on selected segments | Andromeda, October 20262.4% more conversions on Facebook since 2025 |
| Hardware Meta names | Andromeda, December 2024NVIDIA Grace Hopper and Meta’s own MTIA chips | Andromeda, October 2026NVIDIA GH200 Grace Hopper, with Grace Blackwell work under way |
One point about timing. The post describes a system that is in production, not a launch on 28 October. A week earlier, on Meta’s third-quarter results call on 21 October 2026, chief financial officer Susan Li had already credited Andromeda improvements during the quarter, including greater model complexity and throughput, with a 1.4% rise in Facebook conversions. So 28 October is the date Meta explained the system, not the date your delivery changed.
How does Andromeda work with GEM and GenRecs?
Andromeda shortlists ads, GEM teaches the models that rank them, and GenRecs provides a shared way to describe ads for ranking and for a parallel route into the shortlist. Meta’s engineers published a post on each within sixteen days in October 2026, which is why the three names now travel together.
If the table extends beyond the screen, scroll sideways to view all columns.
| System | Its job | What Meta reports |
|---|---|---|
| Andromeda | Its jobRetrieval: scores eligible ads and passes on a shortlist | What Meta reports2.4% more conversions on Facebook since 2025 |
| GEM, the Generative Ads Model | Its jobA very large foundation model that passes what it learns to the smaller models ranking ads in real time | What Meta reportsUp to 18.9% more conversions and 12% higher click-through rate across Meta’s ad surfaces and objectives |
| GenRecs Tokenization | Its jobDescribes each ad with shared tokens for what it shows and says, in place of a separate ID per ad | What Meta reports0.35% more conversions across surfaces when paired with GEM and Meta’s Adaptive Ranking Model |
Two details are often reported wrongly. GEM does not choose your ads. Meta says the model is too expensive to run inside a live ad request, so its knowledge is transferred to the models that do the serving. The 2026 generation uses one central foundation model that learns across Facebook, Instagram and Reels. Specialist models still rank ads in real time.
GenRecs can help new ads with little delivery history. Because an ad is described by what it contains, a new one can borrow what the system has learned from similar ads before it has any history of its own. Matt Steiner, Meta’s vice president of ads infrastructure, ranking and AI foundations, told AdExchanger that this shortens the learning period and does not remove it.
GenRecs also supplies a second route into the shortlist. AdExchanger reports that a small generative model predicts suitable ads from a person’s engagement with ads and ordinary posts, and that its picks are added to Andromeda’s. On the third-quarter call, Susan Li said this generative retrieval launched on Facebook Feed in the quarter and lifted conversions by nearly 2% across the segments it targeted.
What do Meta’s numbers show, and what do they leave out?
Meta reports a large gain in retrieval speed and a 2.4% rise in Facebook conversions since 2025. It gives no baseline or test method for that conversion figure. The claim that matters to advertisers is a single sentence in the engineering post: “Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook.”
Meta has published at least six other percentages about its ad systems since July, and they are easy to confuse.
If the table extends beyond the screen, scroll sideways to view all columns.
| Figure | What it covers | Where Meta said it |
|---|---|---|
| 2.4% more conversions on Facebook | What it coversAndromeda, since 2025 | Where Meta said itAndromeda engineering post, 28 October 2026 |
| 1.4% more conversions on Facebook | What it coversAndromeda improvements in the third quarter of 2026 | Where Meta said itThird-quarter results call, 21 October 2026 |
| 0.24% more conversions on Instagram Reels | What it coversAndromeda improvements in the third quarter of 2026 | Where Meta said itAdvertiser summary, 28 October 2026 |
| Nearly 2% more conversions in targeted segments | What it coversGenerative retrieval on Facebook Feed, third quarter of 2026 | Where Meta said itThird-quarter results call |
| 15.7% more conversions and 8.3% more ad clicks on Facebook | What it coversImprovements to how Meta models user behaviour, combined with GEM ranking and sequence learning, second quarter of 2026 | Where Meta said itSecond-quarter results call, 29 July 2026 |
| Up to 18.9% more conversions and 12% higher click-through rate | What it coversGEM, across Meta’s ad surfaces and objectives | Where Meta said itGEM engineering post, 13 October 2026, and advertiser summary, 28 October 2026 |
| 0.35% more conversions | What it coversGenRecs Tokenization with GEM and the Adaptive Ranking Model, measured in February 2026 | Where Meta said itGenRecs engineering post, 19 October 2026 |
These figures cannot be added together. They cover different periods, different surfaces and different systems, and Meta does not say how they overlap.
The 2.4% claim leaves out most of what you would need to check it. PPC Land and Relevant Audience list the gaps between them:
- no start or end date, baseline or test design;
- no like-for-like Instagram figure for the period covered by the 2.4% claim, although Meta’s advertiser summary separately reports a 0.24% lift on Instagram Reels from Q3 2026 Andromeda improvements;
- no statement on whether the third quarter’s 1.4% is part of the 2.4%;
- no split between the new model and the new machinery.
The engineering figures need care as well. Meta calls this a 100-fold increase. It gives the 2024 rate per GPU card and the 2026 production rate per GPU server, without showing how it calculated the comparison or specifying the server’s GPU count. The post also reports more than 10 billion ads per second for the scoring step alone on one NVIDIA GH200 GPU. That step combines model and system work. The full production pipeline, including targeting checks and shortlisting, processes more than one billion ads per second on a single GPU server. And where the 2024 post gave a recall number, the 2026 post says only that quality approaches ranking level.
Meta’s reported Facebook lift cannot tell you how much your account gained. Meta also says it benefits from lower computing costs. The post itself says each GPU now evaluates more ads at lower cost. On the third-quarter call, Susan Li said Meta’s global average price per ad rose 14% year on year. She attributed the increase to better ad performance and economic conditions, partly offset by strong impression growth on lower-monetising surfaces and in lower-monetising regions. To judge whether your campaigns are improving, track your own cost per customer.
What does Andromeda 2.0 not change?
The Andromeda upgrade requires no change to your Ads Manager settings. Most of the alarming advice in circulation rests on things Meta has not said.
There is no switch and no migration. In its 28 October summary for advertisers, Meta says the upgrades apply automatically, with no setting to enable, new campaign type to adopt or migration to plan.
Your targeting rules still apply. The engineering post describes a step that enforces advertisers’ targeting rules across tens of categories before ads are scored, and says the rebuild kept those rules exactly as they were. In Advantage+ Audience, settings such as minimum age and excluded custom audiences can limit who is eligible. Check which settings are firm controls and whether location expansion is on. Scoring every eligible ad does not mean showing your ad to everyone.
There is no published similarity threshold. Many guides state that Andromeda assigns each ad an “entity ID”, or suppresses ads above a creative similarity score of 40% or 60%. None of Meta’s four engineering posts on Andromeda, GEM and GenRecs contains either term. Solid Labs traced the 60% figure to a tool vendor’s guide and the entity ID idea to a partner training deck.
What Meta has published is narrower. In guidance published on 16 December 2025, Meta says its ads system groups ads with similar visuals and themes and shares learning and delivery optimisations across those variations. Ads Manager also has a creative diversity column that rates an ad set Low, Medium or High. Meta labels the metric as estimated and in development. It applies thresholds to turn the similarity score into a rating, and its Help Centre gives no numerical cut-offs. Akvile DeFazio, writing for Search Engine Land, found accounts she considered varied still rated Low.
There is no required number of ads. In November 2024, Meta recommended at least 20 diversified ads for Advantage+ shopping campaigns. Its December 2025 guidance recommends a broad range of assets across text, image and video without setting a fixed count. Nothing in the October posts sets a quota.
What should UK advertisers do about Andromeda 2.0?
Very little in the account, and rather more in the creative brief. Meta’s posts describe one system for Facebook and Instagram and draw no distinction between countries, so these steps apply to a UK account as they would to any other.
- Leave the settings and the structure alone. There is nothing to enable, and nothing in Meta’s posts supports rebuilding campaigns because a new label is going round.
- Do not date your reports from 28 October. Meta reported Andromeda improvements in the third quarter and described the rebuilt system as already in production on 28 October. A before-and-after comparison around the publication date can show whether your results changed. On its own, it cannot establish that Andromeda caused the change.
- Brief ideas, not variants. This is our inference from the design, and it matches Meta’s own guidance on varied creative. The new model keeps several separate pictures of each person’s interests and reads each ad by its content. Ten edits of one idea may give Meta similar options. Meta says visually and thematically similar ads can share learning and delivery optimisations. Vary the audience addressed, the problem solved, the offer and the format.
- Check your targeting rules. Firm audience controls apply before scoring. Check which exclusions are active and whether location expansion is on.
- Keep the conversion signal clean. The conversion events you send back help Meta measure results and optimise delivery for your chosen outcome. Our Meta Ads UK guide covers the pixel, the Conversions API and the consent limits that apply here.
- Give new ads time. GenRecs helps a new ad start with some borrowed knowledge. Meta’s own executive says the learning period is shorter, not gone, so do not judge an ad on its first few days.
- Treat the creative diversity column as a prompt. A Low rating is a reason to look at your ads side by side. It is an estimate still in development, not a verdict on delivery.
- Judge the result on your own numbers. Compare cost per customer over several weeks, and run a holdout test if your budget allows. Meta reports results for different apps, placements and groups of users.
What has Meta not said yet?
Meta has explained how the system is built and very little about how it was measured or rolled out. Five gaps matter to advertisers:
- How the 2.4% Facebook lift since 2025 was measured, and what Instagram’s lift was over the same period.
- When the 2026 system took over. The post says it is in production and gives no dates, regions or placements.
- How the two retrieval routes combine. Meta names Facebook Feed for generative retrieval and has not said where else it runs or how its picks are weighed against Andromeda’s.
- How grouping of similar ads interacts with scoring every ad. The engineering post does not mention grouping at all.
- How much further it goes. Meta says it expects more gains as further optimisations reach production, and that work on NVIDIA’s Grace Blackwell hardware is under way.
Until Meta fills those gaps, the sound response to Andromeda 2.0 is the unglamorous one: distinct ideas, clean signals and your own measurement.