Meta Andromeda: The Complete Guide for Advertisers (2026)

If your Meta ads changed behaviour over the last two years, with winning ad sets fading faster, broad audiences beating carefully built interest stacks and more of your budget flowing to one or two ads, there is a good chance you have been feeling Andromeda. It is not a setting you switch on. It sits underneath every campaign on Facebook and Instagram and decides which ads get a chance to compete at all.

There is also a lot of noise about it. Some posts treat Andromeda as a mysterious new algorithm that punishes old tactics; others promise a magic number of creatives. This guide sticks to what Meta has published in its engineering blog and earnings calls, marks clearly where advice comes from practitioners rather than Meta, and turns it into a plan you can act on.

It is the hub of our Andromeda series. Each section links to a deeper article: the update timeline, creative diversification, account structure, how many ads per ad set, Andromeda vs GEM vs Lattice, creative testing, broad targeting, Meta ads for B2B SaaS and fixing a performance drop.

What is Meta Andromeda?

Andromeda is the name Meta gave to its personalised ads retrieval engine. Meta introduced it on its engineering blog on 2 December 2024, describing it as a step change in how the company picks ads for people across its apps.

To understand why it matters, you need the shape of Meta's ad delivery system. Every time someone opens Facebook or Instagram and there is room for an ad, Meta runs an auction. But it cannot run its most expensive prediction models on every ad in the system for every person; there are far too many. So delivery happens in stages:

1. Eligibility. Your targeting rules, budget, schedule and policy status decide whether an ad can be shown to this person at all.
2. Retrieval (Andromeda). From all eligible ads, which Meta says number in the tens of millions, a fast model picks a shortlist of a few thousand that are most likely to be relevant to this person right now.
3. Ranking. Larger, slower models predict how likely this person is to take the action each advertiser wants, and estimate ad quality. Meta's ranking side is where systems like Lattice and the knowledge from GEM come in (see Andromeda vs GEM vs Lattice).
4. Auction. The total value of each ad, combining your bid, the predicted action rate and ad quality, decides which ad wins the slot.

Andromeda is step 2. It does not decide the winner. It decides who gets to compete. If your ad never makes the shortlist for a person, your bid, your landing page and your offer never get the chance to matter for that impression.

What Meta has actually said about Andromeda

Because so much of the commentary is speculation, here are the points Meta itself has published, with dates. We go through each in more detail in the Meta Andromeda update explained.

• 2 December 2024, Engineering at Meta. Andromeda introduced as the retrieval engine behind Advantage+ automation, running on NVIDIA Grace Hopper Superchips and Meta's own MTIA chips. Meta reported a 10,000x increase in model capacity for retrieval, a +6% recall improvement and a +8% ads quality improvement on selected segments. It also described a hierarchical index of ads trained together with the retrieval model.
• April 2025, Meta for Business. Meta's guidance shifted to calling creative diversification the main lever for reaching relevant audiences, rather than niche targeting.
• July 2025. Meta reported extending Andromeda to Facebook Reels, with close to 4% more conversions across Facebook mobile Feed and Reels.
• December 2025. Meta clarified that ads which look or feel alike are treated as variations of the same creative, so near-duplicates do not count as diversity.
• 21 October 2026, Q3 earnings. Meta credited Andromeda improvements with a 1.4% increase in Facebook conversions, and said its new generative retrieval model on Facebook Feed drove nearly 2% more conversions in the segments where it launched.
• 28 October 2026, Engineering at Meta. A second-generation Andromeda: over one billion ads scored per second on a single GPU server, which Meta calls about 100x the 2024 system, at around 300x fewer inference operations than the 2024 retrieval model. Meta said Andromeda has delivered a 2.4% increase in Facebook conversions since 2025.

Two cautions. First, these are Meta's own figures, reported as platform-wide averages without a published method, so they tell you the direction of travel, not what will happen in your account. Second, none of them describe a new button in Ads Manager. Andromeda works whether you think about it or not; the question is whether your campaigns give it good material.

Why Andromeda changed how Meta ads work

Older retrieval systems relied heavily on simpler signals: which audiences you selected, which interests a person had, and broad engagement patterns. Andromeda can afford far bigger models at this early stage, which means it can use much richer information about each ad and each person to decide what is relevant. Three practical consequences follow.

1. Your creative is now a targeting signal

Andromeda learns what each ad is about and who responds to it. A video of a founder explaining a pricing problem and a static image showing a product screenshot reach different people, even inside the same broad ad set. That is why Meta now says creative diversification, not tighter interest targeting, is the best lever for finding new customers. Read the full playbook in creative diversification for Meta ads.

2. Near-duplicates waste slots

If you upload ten versions of the same video with different headlines or button colours, Meta has said it treats them as variations of one creative. They compete for the same people instead of opening new pockets of audience. Volume only helps when the ads are genuinely different in concept, format, angle or persona.

3. Fragmented accounts starve the system

Many small ad sets, each with a little budget and a narrow audience, give the system less data per decision and split your learning. Andromeda tends to work best with fewer, broader ad sets that each hold a wide range of creative. We cover the structures that work in Meta ads account structure for 2026.

How to plan campaigns around Andromeda

Here is the approach we use with clients, in the order we would do it. Each step links to a deeper guide.

Step 1: Get your conversion signal right

Andromeda and the ranking models learn from the events you send back. If you optimise for a weak signal, such as page views or cheap form fills, the system will happily find more people who do that and nothing else. Set up the Conversions API alongside the pixel, deduplicate events, and where you can, send deeper events like qualified leads, trials that activate or purchases. This one change often matters more than anything you do with creative.

Step 2: Simplify the structure

For most advertisers, one prospecting campaign per goal with one to three broad ad sets is enough. Use Advantage+ placements unless you have a strong reason not to. Keep retargeting small or fold it into prospecting if your audience is limited. Details and example structures are in Meta ads account structure.

Step 3: Build a diversified creative mix

Map your ads to different buyer motivations, personas and formats: a problem-led founder video, a product demo, a customer story, a comparison, a static with a single bold claim, a carousel of use cases. Meta recommended at least 20 diversified ads for Advantage+ shopping campaigns in late 2024, but for most accounts the right number depends on budget. Our rule of thumb and the maths behind it are in how many ads per ad set.

Step 4: Go broad, and use targeting as a guardrail

Narrow interests now tend to limit Andromeda more than they help it. Use location, age limits where legally required, and exclusions as guardrails, and let creative do the work of finding the right people. When broad does not work, it is usually a sign that the creative or the conversion signal is unclear. See broad targeting on Meta.

Step 5: Test concepts on a schedule

Replace the habit of testing tiny variations with a regular rhythm of new concepts. Put a share of budget into testing (Meta suggested 20 to 30% in its late-2024 guidance for Advantage+ shopping), judge concepts on cost per real result, and move winners into your main ad set. The full framework is in Meta ads creative testing.

Step 6: Diagnose before you panic

Performance on Meta moves for many reasons: seasonality, auction pressure, creative fatigue, tracking breaks and changes on Meta's side. Before you rebuild your account, work through the checks in why your Meta ads stopped performing.

Andromeda for different kinds of businesses

Ecommerce and DTC

Ecommerce brands felt Andromeda first, because they often run large catalogues and Advantage+ sales campaigns. The biggest wins usually come from diversifying beyond product shots: founder stories, use-in-context video, reviews, comparisons and bundles. For a deeper look at online retail, read our ecommerce SaaS marketing guide.

B2B and SaaS

B2B teams often assume Meta cannot work because their buyers are on LinkedIn. With Andromeda, Meta can find job-specific interest from the content of your ads, provided you feed it a signal deeper than a form fill. We cover how in Meta ads for B2B SaaS, and compare channels in our SaaS paid ads guide.

Local and service businesses

Smaller budgets cannot support dozens of ads. The principle is the same, though: three to five genuinely different ads beat ten copies of one. Spend your effort on distinct offers and angles, not on variations.

Common myths about Meta Andromeda

• "Andromeda is a new algorithm that replaced the old one." It is one stage, retrieval, inside a larger system. Ranking and the auction still decide who wins.
• "You need 50 ads per ad set." Meta has not published an Andromeda-specific number. More genuinely different ads help, up to the point your budget can test them.
• "Targeting no longer matters." Meta's October 2026 post confirms that targeting rules are still enforced before scoring. Targeting now acts as a boundary, not a precision tool.
• "Andromeda punishes small advertisers." It rewards clear signals and varied creative. Small accounts can provide both; they just need to be more selective.
• "Andromeda can be switched off." There is no setting. It runs for every advertiser.

What comes next

Meta has said its retrieval pipeline is not yet at its ceiling. In 2026 it began moving generative models into retrieval, starting with Facebook Feed, and its ranking models keep learning from GEM, its large ads foundation model. Expect the direction to continue: more automation, more weight on creative, and fewer manual levers. The advertisers who do well will be those with a clear message, a steady supply of distinct creative and clean conversion data.

If you want help putting this into practice, our paid ads service covers Meta strategy, creative and measurement, and our performance marketing team ties it to pipeline. You can see what that costs on our pricing page.

Frequently asked questions