Meta Andromeda Explained: What It Is, What Changed and Why It's Nothing New

Meta Andromeda is Meta's machine learning system for ad retrieval, the first stage of ad delivery. It narrows tens of millions of ads to a few thousand for each person. Meta deployed it in the second half of 2024 and described it in December 2024. It personalises ad selection, which rewards varied creative, and the principles that worked before still work.

In autumn 2025, Andromeda suddenly became the explanation for every drop in ROAS, and a lot of advertisers were told that it changed everything they knew about Meta ads. I've managed $35M+ in Meta ad spend across 50+ brands, including foodspring, Asana Rebel and Kitchn.io, over about ten years in performance marketing. In this article I explain what Andromeda is in Meta's own words, what it changed, and what I would do with your account now.

I'm Victoria Alenich, a Meta ads consultant, and I teach the same principles you'll read here in my Zero to Ads course.

Last updated: 28 September 2026.

What Is Meta Andromeda?

Meta introduced Andromeda in a post on its engineering blog called Meta Andromeda: Supercharging Advantage+ automation, published on 2 December 2024. In Meta's words, "Andromeda is Meta's proprietary machine learning (ML) system design for retrieval in ad recommendation." Retrieval is the very first step of ad delivery. Meta describes it as "selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates."

In plain English, picture a warehouse holding every ad that is running right now. Meta can't evaluate all of them for one person in a few milliseconds, so Andromeda first pulls a shortlist of a few thousand relevant ads. Only those move on to ranking and the auction.

The post explains that Andromeda runs on "NVIDIA Grace Hopper Superchip and Meta Training and Inference Accelerator (MTIA) hardware". It reports a "+6% recall improvement to the retrieval system, delivering +8% ads quality improvement on selected segments." Meta also mentions "a meaningful increase of model capacity (10,000x) for enhanced personalization". The number of queries the system handles per second went up "by over 3x".

Meta also hints at why it needed better retrieval. "More than a million advertisers used our generative AI (GenAI) tools to create more than 15 million ads in a month," the post notes.

When Was Meta Andromeda Released?

Meta deployed Andromeda in the second half of 2024. On the Q4 2024 earnings call on 29 January 2025, CFO Susan Li said: "In the second half of 2024, we introduced an innovative new machine learning system in partnership with Nvidia, called Andromeda." She described "a 10,000x increase in the complexity of models we use for ads retrieval". She added, "This has driven an 8% increase in the quality of ads that people see on objectives we've tested."

Andromeda is one step in a much longer and more gradual evolution, as the dates show.

Meta kept improving the system after launch. On the Q3 2025 earnings call, Meta said it "significantly improved performance of Andromeda in Q3 by combining models across retrieval and early stage ranking into a single model." That drove "a 14% increase in ads quality on Facebook surfaces." On the Q4 2025 earnings call in January 2026, Meta said Andromeda "can now run on Nvidia, AMD, and MTIA". It also said it had managed to "nearly triple Andromeda's compute efficiency."

The automation story itself is older, since Meta's Advantage+ shopping campaigns announcement from August 2022 already let the system test creative combinations for you.

How Andromeda Fits Into How Meta Picks Your Ads

I like to describe the Meta algorithm as a silent super employee. It's the best media buyer in the world, working for you 24/7 for free, "but this employee can't talk." You communicate with it through your optimization event, meaning your tracking, and through your creative. Everything I teach comes back to working with the algorithm instead of fighting it.

As I tell my students, "No one really knows how it works, not even all Meta employees fully understand it." You can't game it, so the best you can do is understand its broad shape and give it what it needs. That broad shape has three stages, and Andromeda runs the first one.

After Andromeda builds the shortlist, ranking models decide how promising each candidate is. Meta's GEM post from November 2025 describes its generative ads ranking model. Meta credits it with a "5% increase in ad conversions on Instagram and a 3% increase in ad conversions on Facebook Feed in Q2" of 2025.

Then comes the ad auction, where total value equals your bid multiplied by your estimated action rate, plus ad quality and value to the user. Meta protects the experience of the people scrolling, so the biggest bid alone rarely wins. You're also competing with every advertiser who wants that same person, including brands that sell something completely different. The way to win more auctions is a higher estimated action rate, through the right event and better creative.

Why Did Everyone Suddenly Start Talking About Andromeda?

Search interest in Andromeda was close to zero before March 2025 and started climbing in August 2025. It peaked in the week of 26 October to 1 November 2025, the same week Meta talked about Andromeda on its Q3 earnings call. By late 2026 it had settled to a fraction of that peak.

So the hype arrived roughly 9 to 11 months after Meta first described Andromeda publicly, and more than a year after it went live. If your account ran through 2024, Andromeda was already choosing your ads long before anyone told you to panic about it.

What worries me is how some of the conversation turned into fear. Claims that Andromeda changed everything were quickly followed by an Andromeda fix, course or playbook for sale. In my experience, anyone who scares advertisers with a sudden algorithm apocalypse and then sells the cure is probably not a very trustworthy source.

These are the questions I'd ask of any advice about Andromeda, or about the next big update after it.

• Does it link to what Meta actually said, like the engineering post or the earnings calls, or only to someone's opinion?
• Would the same advice have worked two years ago? If it would, the principle is sound and the Andromeda label is just packaging.
• Is the message built on urgency, countdowns and the idea that everything you knew is now wrong?
• Are the numbers used the way Meta used them, or has a figure like 10,000x been stretched into something Meta never said?

For a sober outside view, AdExchanger's piece on what the Andromeda update actually changes and what it doesn't from January 2026 is a good read.

Myths About Meta Andromeda

Here's how the most repeated claims compare with what Meta has actually published.

What Actually Changed for Advertisers?

Some things really have shifted over the last few years, although gradually, and they point in the same direction I've been teaching for years.

Creative diversity matters more

In April 2025, Meta published The Creative Advantage: Unlocking the Power of Diversification with Meta Andromeda. It says that "the focus has shifted from niche targeting to creative diversification as the best lever to find the most relevant audiences." A later article, Demystifying Creative Diversification, asks advertisers to "build a broad portfolio of assets that take more creative liberty across text, image, and video, reflecting understanding of multiple customer personas."

That makes sense once you picture retrieval. A personalised shortlist can only pick your ad for someone if you have an ad that fits them. Your creative acts as targeting, since different hooks attract different customers and each one tells the algorithm who is likely to convert.

In my experience, improving targeting might get you 10 to 15%, bidding 10 to 20% and structure 20 to 25%. The right creative can move results 3x, 4x, 5x or more. One of my skincare clients had a single before and after video that drove 80% of their sales. It did that at a third of the cost of every other ad, with the same audience, product and budget.

Broad audiences and Advantage+ are the default

From Marketing API v23.0, Advantage+ audience is switched on by default for new ad sets, so your targeting inputs act more like suggestions. That's why I call your optimization event "possibly the most powerful targeting tool you have". The broader your targeting, the more your results depend on it. More on this in my guide to Facebook ads targeting.

Simpler structures work better

Meta keeps consolidating on its side too. On the Q3 2025 call it said: "Since introducing Lattice back in 2023… we have now cut the number of ads ranking and recommendation models by approximately one hundred." When the system does more of the matching, splitting your budget across many tiny ad sets mostly takes data away from it. As I put it in my course, "It's better to have a simple, well structured campaign than a complex one that confuses the algorithm."

Creative fatigue comes faster

Back in 2019, a strong ad could often run for a month or two before it wore out. In many accounts I see now, fatigue shows within about a week. That's the most practical change of all, because you need a steady flow of new creative more than a big launch every quarter.

What Should You Do About Meta Andromeda Now?

None of these steps were invented for Andromeda, which is exactly why I trust them. They're the principles I teach in Zero to Ads.

1. Fix your tracking and choose the right optimization event

Your event is how you talk to the silent super employee. Make sure your Pixel and Conversions API send clean data, and optimize for the action that actually makes you money. Broad targeting with a valuable event often beats narrow targeting with a weak one.

2. Simplify your campaign structure

Consolidate overlapping ad sets so each one gets enough data to learn from. One of my clients consolidated 10 ad sets into 3 and cut their cost per purchase by 30% within two weeks. My guide on how to scale Facebook ads shows how I build structures that grow without breaking.

3. Test genuinely different concepts inside CBO

I test new creative inside a campaign with campaign budget optimization, because slow classic A/B tests hold you back. The idea is to "feed the algorithm good options and let it optimize." Your concepts should differ in angle, hook, format and the customer they speak to. For a B2B software client, raw CEO videos brought the cost per lead down from €280 to €45.

4. Keep a steady creative testing cadence

With fatigue arriving faster, add a small batch of new concepts every week or two and let the winners keep running while the new ones compete.

5. Judge on trends, and respect the optimization zone

Judge performance on trends over at least two weeks, since one or two days tell you very little. If you're within 10 to 20% of your target, you're in what I call the optimization zone, where you tweak and keep things running. Budget changes should stay within 20% a day.

Is Andromeda Why My ROAS Dropped?

Changes on Meta's side can play a part. In the accounts I review, though, a drop in ROAS almost always has a more ordinary cause that you can find and fix. Before you blame Andromeda, work through these checks in order.

If CPMs are the issue, my article on why your Facebook ads CPM is so high explains what drives them. For a wider checklist, see why your Facebook ads are not working. You can also compare your numbers with current benchmarks in how much Facebook ads cost.

When an account has drifted too far for small fixes, I "reset and rebuild" with broader targeting and new concepts. As I say in my course, "there's no shame in starting over." If you want a second pair of eyes first, a Facebook ads audit is the fastest way to find out what's going on.

Frequently Asked Questions About Meta Andromeda

The Principles Stay the Same

The accounts that do well year after year have clean data, simple structures and a steady stream of creative that speaks to real customers. As I tell my students, "That shiny new feature everyone's talking about? It'll be old news in six months. But the principles of working with the algorithm? That's what stays constant."

Meta ads isn't rocket science once you know what the algorithm is listening for. If you'd like me to look at your account with you, read more about how I work as a Facebook ads consultant.