The Meta Andromeda Update Explained: Every Change From 2024 to 2026

"The Andromeda update" gets blamed for a lot: rising CPMs, falling ROAS, campaigns that stopped scaling. Some of that is fair and some is not. The best way to separate the two is to look at what Meta actually changed, and when.

This article is a dated timeline built from Meta's own engineering posts, business guidance and earnings calls. For the big picture, start with our complete guide to Meta Andromeda.

Before Andromeda: why retrieval became the bottleneck

Meta's ads system has to choose from an enormous pool of ads in milliseconds. The ranking models that predict conversions are expensive, so they can only look at a small shortlist. That shortlist comes from retrieval.

In 2023 and 2024 two things made the old approach strain. Advantage+ automation encouraged advertisers to run broader campaigns with more creative, and Meta's generative AI tools made it cheap to produce new versions of ads. Meta said in December 2024 that more than a million advertisers had used its generative AI tools to create more than 15 million ads in a single month. The pool of eligible ads exploded, and a retrieval system that could not consider most of it would miss good matches. Andromeda was Meta's answer.

December 2024: Andromeda launches

On 2 December 2024 Meta's engineering blog introduced Andromeda as its personalised ads retrieval engine. The main points:

• Hardware co-design. Built to run on NVIDIA Grace Hopper Superchips and Meta's own MTIA accelerators, with feature extraction moved from CPUs to GPUs. Meta reported more than 100x better feature extraction latency and throughput, and over 3x more inference queries per second end to end.
• Much bigger models. A 10,000x increase in the capacity of retrieval models, with inference cost growing more slowly than capacity.
• A smarter index. Ads organised into a hierarchical index that is trained jointly with the retrieval model, rather than a fixed lookup.
• Results. A +6% improvement in recall (finding the relevant ads) and +8% in ads quality on selected segments.

The same post connected Andromeda to Advantage+: advertisers who had not used Advantage+ creative before and switched on its AI-driven features saw, according to Meta, a 22% increase in return on ad spend. That figure is about Advantage+ adoption, not Andromeda alone, but it shows how Meta sees the two working together.

What it meant for advertisers

Nothing changed in Ads Manager. What changed was that more of your ads could be matched to more people on the strength of what they contained. Advertisers with varied creative started to see different ads win for different audiences inside one ad set; advertisers with one hero ad saw that ad fatigue faster.

2025: creative diversification becomes the official advice

Spring 2025: the guidance changes

In March 2025 Meta explained that Andromeda works behind the scenes as advertisers upload more creative, and in April 2025 it published guidance describing creative diversification as the best lever for finding relevant audiences, replacing the old emphasis on niche targeting. Practitioners also noticed that the long-standing rule of thumb of around six ads per ad set quietly disappeared from Meta's advice around this time.

July 2025: Andromeda reaches Facebook Reels

Meta reported extending Andromeda to Facebook Reels, with nearly 4% more conversions across Facebook mobile Feed and Reels. For advertisers, this meant vertical video built for Reels now benefited from the same matching.

November 2025: GEM is detailed

On 10 November 2025 Meta's engineering blog described GEM, its Generative Ads Recommendation Model, as the largest foundation model for recommendation in the industry. GEM teaches Meta's other ad models through knowledge transfer. Meta said it had lifted ad conversions by 5% on Instagram and 3% on Facebook Feed in Q2 2025. GEM is not Andromeda, but the two are often confused; see Andromeda vs GEM vs Lattice.

December 2025: similar ads count as one

Meta clarified that ads which look or feel alike are treated as variations of the same creative. This was the clearest signal yet that uploading many near-identical versions does not buy you more reach. It also came with a creative diversity indicator that rates ad sets as high, medium or low, which Meta describes as an estimate still in development.

2026: generative retrieval and a rebuilt engine

July 2026: the first generative model in retrieval

On its Q2 2026 earnings call on 29 July 2026, Meta said it had deployed its first generative model into its ads retrieval system. Generative retrieval works differently from classic scoring: instead of scoring each ad against a person, the model learns to generate the identifiers of ads a person is likely to engage with.

21 October 2026: Q3 results

On the Q3 2026 call Meta credited Andromeda improvements with a 1.4% increase in Facebook conversions, and said the Meta Generative Recommender, launched as generative retrieval on Facebook Feed, drove a nearly 2% increase in conversions across the segments where it was used.

28 October 2026: second-generation Andromeda

A week later Meta's engineering blog described a rebuilt Andromeda. The key changes:

• A mixture of experts. Instead of one model of the user, a set of specialised "expert" models, with an attention-based gating network that routes each candidate ad to the most relevant experts. Meta says this approaches ranking-level quality at about 300x fewer inference operations than the 2024 retrieval model.
• Everything on one GPU. Index matching, scoring and picking the top results now run on the same GPU, so only the shortlist travels over the network. Meta says this cut network traffic by about 100x.
• Speed. Over one billion ads scored per second on a single GPU server, which Meta calls about 100x the 2024 rate.
• Targeting still applies. Index matching enforces each advertiser's targeting rules before scoring, so "every eligible ad" means every ad whose targeting includes this person.
• Results. Meta says Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025.

Meta also said the pipeline is not yet at its ceiling, which tells you to expect further changes.

How to read Meta's numbers

Every figure above comes from Meta, and Meta has good reasons to present its systems well. That does not make the numbers wrong, but it does mean you should read them carefully:

• They are averages across the platform. A 2.4% lift across Facebook says nothing about whether your account will see more or less.
• Methods are not published. Meta does not share the baseline, time window or test design behind most of these claims.
• Figures overlap. The 1.4% Q3 figure and the 2.4% "since 2025" figure may cover different periods or changes, so do not add them together.
• Better for the platform is not always better for you. More conversions overall can come with higher prices per ad. Meta also reported on the Q3 2026 call that the average price per ad rose 14%.

What each update means for your account

Across all these changes, the practical advice has stayed consistent and has only become stronger:

1. Feed it variety. Each update made the system better at matching different ads to different people. That only helps if you have different ads. Start with creative diversification.
2. Give it room. Broad targeting and consolidated ad sets let retrieval consider more people for each ad. See broad targeting and account structure.
3. Send better signals. The system optimises for what you report. Report outcomes that matter, using the Conversions API and offline events. Our performance marketing system explains the setup.
4. Test steadily. New concepts on a schedule beat big rebuilds when results dip. See Meta ads creative testing.
5. Watch costs, not just conversions. If Meta's average price per ad is rising, efficiency comes from better conversion rates after the click too. Our conversion rate optimization guide covers the landing page side.

Frequently asked questions