Meta Andromeda: What It Is and What It Changed in Ads

Meta Andromeda is the name behind most of the advice Facebook and Instagram advertisers have been given since 2025: fewer ad sets, broader targeting and many more creatives. Much of that advice comes from agencies rather than from Meta, and it is easy to lose track of what Meta has actually said. This guide separates Meta's own statements about Andromeda from the commentary around it, and explains how to check the results in your own store's data.

The short answer: Andromeda decides which ads are considered, and Arktis checks what reached your store

Meta Andromeda is the retrieval stage of Meta's ad delivery system: in Meta's words, it narrows tens of millions of candidate ads to a few thousand before ranking models such as GEM pick what a person sees. Meta described it publicly in December 2024, says it improved ads quality by 8 percent, and has kept extending it since, most recently reporting a 1.4 percent increase in conversions on Facebook from Andromeda improvements in the third quarter of 2026. What it changed for advertisers is the creative advice: Meta now says creative diversification, not niche targeting, is the best lever for finding relevant audiences, and that ads which look or feel alike are treated as variations of the same creative.

Arktis does not change how Meta delivers ads and does not replace Ads Manager. What it is the clear winner for is the outside check on any restructure: it records fbclid when a visit arrives with it, matches Shopify orders back to a visit, counts conversions per ad platform and per UTM campaign, and syncs Meta spend per campaign once a day. Plans are $49, $149 and $349 a month, with a 7-day free trial on Growth.

Key takeaways

Andromeda is not a setting and there is nothing to switch on. It is the part of Meta's system that decides which ads are even considered for a person, and Meta has kept changing it since 2024. Meta's practical advice that follows from it is narrow: make genuinely different creatives for different customers rather than many small variations of one ad, and let automation choose who sees which. Meta's performance figures for Andromeda are measured by Meta across its whole system, so they say nothing about your account. Judge any restructure by orders your own store can match to Meta visits.

What Meta Andromeda is

Meta's engineering post of 2 December 2024 describes Andromeda as its proprietary machine learning system for retrieval in ad recommendation. Retrieval is the first step: it selects a few thousand relevant ads from tens of millions of candidates, and larger ranking models then decide which ads are shown. The post says retrieval handles three orders of magnitude more ads than later stages under tight time limits, and that the number of ads was expected to grow significantly as advertisers adopted generative AI tools.

Meta's answer was a deep neural network designed for Nvidia's Grace Hopper hardware, with what it calls a 10,000 times increase in model capacity and a hierarchical index to cope with many more creatives, in a system Meta says was co-designed with its own MTIA chips. Its deployment across Instagram and Facebook, the post says, achieved a 6 percent recall improvement and an 8 percent ads quality improvement on selected segments.

The plain-language version came in a Meta for Business post on 27 March 2025, which placed Andromeda alongside three other systems: GEM, a large ads recommendation model used in ranking; Lattice, a ranking architecture shared across objectives and surfaces; and sequence learning. It says that as businesses upload more creatives to support a diversification strategy, Andromeda works behind the scenes to pick the right creative, and repeats the 8 percent ads quality figure.

What Andromeda changed in ad delivery

According to Meta, the change is in how many ads, and how many versions of each ad, the system can weigh for one person. Meta's December 2024 post says the earlier retrieval systems could apply only limited personalisation and relied on rule-based heuristics; Andromeda cuts components and rule-based logic so the system can be optimised end to end.

Meta has kept changing it. On its third quarter 2025 follow-up call, Meta's vice president of finance, Chad Heaton, said Meta had consolidated retrieval and early-stage ranking into a single model, which drove a 14 percent increase in ads quality on the Facebook surfaces where it was introduced, Feed, Reels and Marketplace, and that Meta planned to roll Andromeda out to Instagram Feed and Reels later that quarter. On 28 January 2026, chief financial officer Susan Li said Andromeda had been extended to run on Nvidia, AMD and MTIA hardware and that its compute efficiency had nearly tripled.

On its 29 July 2026 earnings call, Meta said it had introduced Meta Generative Recommender that quarter, using large language models to reason about ad content and user preferences together, and had deployed the first generative model into its ads retrieval system. On 21 October 2026, Li said generative retrieval launched on Facebook Feed drove a nearly 2 percent increase in conversions across targeted segments, and that improvements to Andromeda raised conversions on Facebook by 1.4 percent in the quarter.

All of these figures are Meta's, measured across its own system with its own definitions of ads quality and conversions. They describe averages over many advertisers, not what any one account should expect.

What Meta tells advertisers to do about creative

Diversify, do not iterate

Meta's post of 22 April 2025 defines creative diversification as making a wide range of ads with different themes, messages and visuals for different audience segments, and says the focus has shifted from niche targeting to creative diversification as the best lever to find the most relevant audiences. Its post of 16 December 2025 draws the line more sharply: creative iteration produces two ads with identical visuals but different text, while diversification produces two distinctly different pieces of creative, such as a Reels ad for one persona and a creator partnership ad for another.

Similar ads count as one creative

The December post also says Meta's system groups ads that share key visual and thematic attributes. When ads look or feel alike, they are treated as variations of the same creative, and learnings and delivery optimisations are shared at the creative level. Meta separates creative fatigue, where people have seen the same ad too often, from creative similarity, where different ads look too alike, and says it is testing creative insights in Ads Manager to flag both.

Simplify the account

Meta's Performance 5 page lists account simplification, automation, creative diversification, data quality and results validation as its five practices. On simplification it advises combining ad sets and minimising changes during the learning phase, and claims that advertisers who keep under 20 percent of spend in the learning phase can lower cost per purchase by as much as 68 percent.

None of the Meta pages above gives a number of creatives to run per ad set or ties a creative count to Andromeda.

What agencies add, and where it goes beyond Meta

Much of what is written about Andromeda is commentary. Jon Loomer, who writes and trains on Meta advertising, reads Meta's posts as a reason to keep ad sets broad and let Andromeda match people to ads, to prefer Advantage+ defaults and to simplify structure. He notes that Meta once advised no more than six ads per ad set and that this guidance disappeared, which he links to Andromeda while saying Meta never stated that. He runs an ad set with 50 ads himself, but warns that more ads do not guarantee results.

That is a reasonable reading, but it is a reading. Meta's own pages say creative diversification is the best lever for finding audiences; they do not say interest targeting has stopped working, and they publish no creative count. Agency guides that give a fixed number of concepts per ad set are offering their own rules of thumb.

How to measure Andromeda-era results outside Meta

Meta's own help centre is careful about comparisons. Its article on attribution settings says results in the Campaign Overview table cannot be compared across ad sets with different attribution models, and adds that if you use external analytics tools you should evaluate performance in those tools. A restructure into fewer, broader ad sets with more creatives changes delivery, and Meta's reported conversions follow Meta's attribution settings, which, under the standard model, count a purchase within 1 or 7 days of a link click, 1 day of an impression, or 1 day of another click on the ad or, for video ads, a play of at least 5 seconds, although Meta says some accounts are still on earlier versions of these settings. Our guide to attribution windows covers those settings.

Three steps keep the comparison honest. Keep the attribution setting unchanged before and after the restructure, so Meta's own numbers are comparable. Give the new structure its own utm_campaign values with the free UTM builder, so visits and orders before and after sit in separate rows. Then compare, for the same weeks, Meta's reported purchases with the orders your store can match to Meta visits; if Meta's count rises while matched orders do not, the gain is in Meta's counting rather than your sales.

In Arktis, an order counts for Meta only when it was matched to a visitor who arrived with fbclid, so the Meta figure is a click-based count. Conversions are shown per ad platform and per UTM campaign, from matched orders and the conversion events the tracker records. The Models view compares five rule-based models, first-touch, last-touch, linear, time-decay and position-based, across ad platforms with lookbacks of 7 to 90 days, which shows whether Meta earns more credit as the first touch or the last. The Revenue page shows revenue net of refunds by the source of the visit each order was matched to.

If you need to know whether Meta ads cause sales at all, a click-based count is the wrong tool; a holdout test is the right one, as our guide to incrementality testing explains.

Where Arktis fits, and where it does not

Andromeda's whole premise is matching individual creatives to individual people, and Arktis shows Meta spend per campaign rather than per ad, with creative detail only where each ad carries its own utm_content value, so the question of which of your 30 creatives worked belongs in Ads Manager and its creative insights. Arktis does not know which shoppers saw an ad without clicking it, so it will never reproduce Meta's view-through conversions, and it does not send events to Meta's Conversions API, so it does not change what Meta's system learns from.

Matching has limits too. Shopify orders arrive by webhook from the day the store connects, with no bulk import of earlier orders. They are matched to a visit on email where the store has identified the visitor, otherwise on the UTM source of a session from the previous 24 hours, which can be another visitor's session. Orders that cannot be matched are kept and appear under Direct on the Revenue page. Arktis identifies visitors per browser and does not join one person across devices, so a shopper who clicks a Meta ad on a phone and buys later on a laptop is usually not credited to Meta.

Check your Meta results against your own orders

To see Meta-matched orders next to your other channels, start the 7-day Growth trial, or compare plans on the pricing page. For related changes, read Meta product browsing, why Facebook Ads and Shopify sales don't match and fbclid tracking.

Sources

Engineering at Meta: Meta Andromeda, the next-gen personalized ads retrieval engine, retrieval stage, model capacity, recall and ads quality figures, published 2 December 2024

Meta for Business: AI innovations in Meta's ad ranking driving advertiser performance, Andromeda, GEM, Lattice and the 8 percent ads quality figure, published 27 March 2025

Meta for Business: The Creative Advantage, unlocking the power of diversification with Meta Andromeda, from niche targeting to creative diversification, published 22 April 2025

Meta for Business: Demystifying creative diversification, iteration against diversification, grouping of similar ads, fatigue and similarity, published 16 December 2025

Meta for Business: Performance 5, the five practices and the learning phase claim, accessed 26 October 2026

Meta third quarter 2025 follow-up call transcript, consolidated retrieval model, 14 percent ads quality and the Instagram rollout plan, call held 29 October 2025, accessed 26 October 2026

Meta fourth quarter 2025 earnings call transcript, Andromeda on Nvidia, AMD and MTIA and compute efficiency, call held 28 January 2026, accessed 26 October 2026

Meta second quarter 2026 earnings call transcript, Meta Generative Recommender and the first generative model in ads retrieval, call held 29 July 2026, accessed 26 October 2026

Meta third quarter 2026 earnings call transcript, 1.4 percent conversion increase from Andromeda on Facebook and generative retrieval on Facebook Feed, call held 21 October 2026, accessed 26 October 2026

Jon Loomer Digital: Meta Andromeda, what it means for your ad strategy, consultant commentary on structure, targeting and ad counts, accessed 26 October 2026