Meta's Andromeda Update, Decoded for Operators: What Actually Changed (and What Didn't)

It's a Tuesday night, the house is quiet, and a founder I'll call you is staring at a dashboard that's gone the wrong colour. ROAS that sat comfortably at 2.4 a month ago is reading 1.6. Nothing in the account changed. Same ads, same budgets, same broad targeting. So you do what everyone does at 11pm with a coffee going cold: you open three browser tabs, and every one of them is screaming the same word at you. Andromeda.

One video says it's killing ROAS across the board and you need to act ASAP. Another calls it the biggest shift since iOS 14. A third tells you, calmly, that nothing has really changed. They cannot all be right, and you've got payroll on Thursday.

So let me do the thing none of those tabs quite did for me when I first read the whitepaper. Translate it into plain English, separate what actually moved from what didn't, and tell you what I'd do with your account on Wednesday morning.

What the whitepaper actually says, minus the jargon

Meta published the Andromeda paper on the 2nd of December 2024. It then rolled out quietly across the year, account by account, Advantage+ first. That timing matters, and I'll come back to it.

The paper is dense, but it's really solving two problems Meta gave itself.

Problem one: too many ads to choose from. Every time someone opens Facebook or Instagram, Meta has to decide, in a sliver of a second, which handful of ads to show them out of an enormous pool of candidates. Think of it as a casting call with millions of auditions and one slot. The old system could only seriously consider a limited shortlist before it ran out of time.

Problem two: not enough time to decide. That sliver of a second is a hard ceiling. Meta calls it a latency constraint. In normal words: the system has to pick the right ad for the right person almost instantly, and doing that well takes a frightening amount of computing power.

Andromeda is the machinery they built to do both at once. The line everyone quotes is a "10,000x" increase in model capacity. I wouldn't treat that as a number you can do anything with, it's a flex about how much bigger the engine is. What it means in practice is the system can now hold far more context about you, the scroller, and sort through far more ad candidates, and still decide in time.

"One-to-one retrieval" in human terms

Here's the part that actually changes how you should think.

The old way matched ads to audiences. You're a 34-year-old woman who likes yoga, here's the yoga-ish ad. Buckets of people, buckets of ads, rough pairing.

Andromeda matches the ad to the individual. Not the bucket you're in, you specifically. Your full history, the things you've lingered on, the styles of content you actually stop scrolling for, all of it. One commentator described it as Meta trying to find the soulmate ad for every single person, and if it had infinite ad options it could find the perfect one. That's a romantic way to put it, but it's the right mental model.

So the question the algorithm is asking has shifted. It used to be "which audience does this ad belong to". Now it's "of all the ads available to me, which single one is this exact human most likely to buy from right now". The more genuinely different ads you give it to choose from, the more often it finds a real match. Give it five near-identical videos and you've handed the casting director five actors who look the same.

The change that's real, and the noise that hid it

This is where I have to be honest, because the panicky tabs and the calm tab are both telling a half-truth.

The mechanical change is real. Meta rebuilt the retrieval engine, and it now rewards creative variety in a way the old system couldn't. That's not marketing fluff, it's in the architecture.

But here's the thing nobody freaking out at 11pm wants to hear: most of the ROAS dip people blamed on Andromeda wasn't Andromeda. It was the calendar.

The update finished rolling out for a lot of accounts right as the back half of the year did its usual seasonal thing, and then Q4 noise sat on top of it. If you've run ads for more than a couple of years you know certain stretches are just soft. People travel, schedules reset, attention drifts, and then the platform itself has the odd delivery wobble. Stack three or four of those on the same fortnight and your account looks like it's been gutted, when really the season changed and the algorithm was mid-migration.

I find it useful to separate the two cleanly. Did Andromeda alter the plumbing? Yes. Did Andromeda personally tank your ROAS in a given week? Almost certainly not on its own. The brands that screenshotted a cliff and captioned it "Andromeda" mostly caught a coincidence and gave it a name.

And naming things is exactly how a panic spreads. A real mechanical change lands, performance dips for unrelated seasonal reasons, and suddenly there's a tidy villain to chase. So everyone reaches for the new fix, feels briefly better, and never checks whether the villain was actually responsible.

What Andromeda didn't break: it just formalised good habits

Here's my actual take, and it's the unglamorous one.

Andromeda didn't introduce a new way to run ads. It put Meta's full weight behind the way the best operators were already running them.

Go back a few years and Meta's own advice was "simplify your account, focus on creative". Then they removed most manual targeting and told us broad was fine. Andromeda is the next step in the same straight line. It's not "creative matters" anymore, because every advertiser already nods along to that. It's "creative variety matters", which is a sharper instruction: stop running the same idea ten ways, and start running ten genuinely different ideas.

That's a best practice I'd have given you in 2022. The difference is now the algorithm enforces it instead of merely preferring it. Which, honestly, I think is good news for any founder willing to do the work, because it means the thing you can actually control, the creative, is the thing that pays off.

To put the upside in perspective: Meta has pointed to roughly an 8% improvement in ad quality for users and a meaningful ROAS lift for advertisers leaning into the Advantage+ tools that ride on this engine. I'd treat any single headline figure with a pinch of salt, but the direction is the point. Feed the system variety and it rewards you. Starve it and it doesn't.

What I'd actually change in your account on Wednesday

Enough theory. If Andromeda is the new reality, here's what it means for the three levers you care about. None of this is exotic.

Budget: stop force-feeding. The fastest way to get punished under this system is to chop your budget into a dozen tiny campaigns, each with a forced daily spend on ads that don't deserve it. You're fighting an engine that's designed to flow money toward whatever's matching best, in real time. Let it. A consolidated structure, with the budget at the campaign level, lets Andromeda do the one thing it's genuinely good at. Twenty active campaigns with little fixed budgets each is the surest way to pay what some people are calling the "Andromeda tax".

Structure: fewer, wider campaigns. Consolidate. One prospecting campaign on a campaign-budget setup, broad, with your creatives competing inside it, beats a sprawl of fragmented ad sets in almost every account I see. Keep separate campaigns only where you have a real reason, a genuinely different margin profile, or an inventory commitment, say you need spend going to both the men's and women's range no matter what. Otherwise, let it consolidate.

Creative: variety, not volume for its own sake. This is the one that matters most, and it's the one people get wrong. "More creative" does not mean the same winning video with fifty slightly different opening frames. It means real variety, different angles, different formats, different people, different reasons-to-buy. A useful discipline I like: roughly 80% of new creative as iterations and variations on what's working, and 20% genuinely new messaging. That keeps the system fed with fresh, distinct options without you reinventing the brand every week.

Notice that not one of those moves is new. If you'd been running broad, consolidated, and creative-led already, Andromeda barely touched you. That's not luck. That's the whole point of what changed.

Where to from here

So the honest answer to your three browser tabs: the plumbing changed, the panic mostly didn't need to, and the fix is the same boring, durable thing it's been for years. Variety in, performance out. Consolidate, go broad, and stop running the same idea ten ways.

If you want to know whether your account actually caught the change or just caught a soft season, that's a question your numbers can answer cleanly, you just have to look at the right window rather than the scary one. A Signal/Noise Audit is the version of that I run: a proper read of your account structure, your creative history, and your unit economics, to tell you where the genuine drag is sitting versus where you're chasing a ghost. No pressure either way. But before you rebuild anything at 11pm on the strength of a video, it's worth knowing which problem you're actually solving.