Meta’s Andromeda Update, Explained

If your Meta ad results have moved a lot over the past year, and you did not change your targeting, the most likely reason is that Meta did. Meta has spent 2025 and 2026 changing how it picks which ad to show which person. It started with a system called Andromeda. Now there is a bigger system called GEM. Meta did not announce these changes the way it announces a new campaign type. They showed up first as technical blog posts, and then as changes in your numbers that nobody explained.

This post explains what actually changed, why your numbers moved, and what to do about it. Most articles on this topic skip that last part.

What Is Meta Andromeda, Actually?

Andromeda is the system that runs the very first step in how Meta picks an ad for you. Out of tens of millions of possible ads, it has to narrow the list down to a few thousand before any of them get properly scored (Meta Engineering, December 2024). This step has to happen in milliseconds, so for years Meta used simple rules to do it fast.

Andromeda replaces those simple rules with deep learning models. Andromeda also sorts ads into layers, a bit like a filing system, so it does not have to check every single ad one by one. And it can spend more computing power on ads that matter more, and less on ads that do not.

In plain terms: Meta used to filter ads with basic rules, then use AI to rank the ones that made it through. Now Meta uses AI at the filtering step too. That one change is why a lot of the old advice about ad structure and creative stopped working as well as it used to.

Here are the numbers Meta shared when it launched Andromeda:

• 10,000 times more model capacity at this filtering step

• Over 100 times faster at pulling the data it needs out of each ad

• 10 times more efficient at running its models, and 3 times more searches per second

• 6% better at finding relevant ads, and 8% better ad quality on the first group of ads it tested this on

• 22% higher ROAS for advertisers using Advantage+ creative, and 7% more conversions for advertisers using Meta’s image-generation tools

The last two numbers matter most for this post. Andromeda does not just rank ads better. It rewards accounts that give it more, and more varied, creative to work with.

From Andromeda to GEM: What’s New in Meta’s Ad System

Andromeda solved the first step: picking which ads are even worth considering. It did not touch the steps after that, where Meta ranks and scores the ads that made it through. That’s what GEM is for.

GEM stands for Generative Ads Recommendation Model. Meta introduced it in November 2025. It is one large model, built the same way large language models are built, and trained on both ad content and how people actually respond to ads (Meta Engineering, November 2025). Meta says GEM led to a 5% rise in ad conversions on Instagram.

In August 2026, Meta published a deeper technical post about GEM. This one was mostly about training GEM faster and cheaper, not about a new feature for advertisers (Meta Engineering, August 2026). Meta doubled how efficient GEM’s training is, and used four times more computing power to train it than it did a year before.

That post is a signal worth paying attention to. Meta is still investing heavily in GEM. It is not treating GEM as finished. If your numbers moved again in the last few months with no clear reason, GEM’s ongoing changes are a more likely cause than most of the theories floating around online right now.

Here is the short version of how the two systems fit together. Andromeda decides which ads are even worth showing to you. GEM decides which of those ads actually wins, and how it gets shown. Both systems get better, and reward you more, the more real variety you give them to work with.

Why Your CTR and ROAS Moved After the Rollout

Three things are true at the same time here, and most articles on this topic only mention one of them.

First, Andromeda rewards variety, not just volume. The old advice was simple: upload more ads. That advice does not work as well anymore. The new goal is to give the system more genuinely different ideas to choose from. Ten ads that are the same idea with different backgrounds do not help much. Ten ads built from different ideas help a lot more.

Second, both systems personalize harder than before. That means results are less steady when an ad or account is new, or has little data yet. Small accounts and new campaigns will see more noise in their numbers while these systems learn, even if the advertiser did nothing wrong.

Third, it got harder to tell targeting problems apart from creative problems. Targeting and creative choices increasingly happen inside the same system now, especially with Advantage+ campaigns. So a drop in CTR or ROAS that looks like a targeting issue might actually be a creative issue, and the other way around. We cover how to actually test for this in a separate post on proving what drives ROAS. The short version: you need a real test, not just a look at your account-level numbers.

Andromeda vs. Advantage+: How They Fit Together

People mix these two up all the time, but they are not the same thing.

Advantage+ is the product you actually choose in Ads Manager. It is the campaign type and automation settings you turn on, like Advantage+ shopping campaigns or Advantage+ creative.

Andromeda is not something you choose. It is the system running underneath every campaign type, whether you use Advantage+ or not. You cannot opt out of it. GEM works the same way. It sits at the ranking step, also running underneath everything.

What this means in practice: you cannot avoid these changes by staying away from Advantage+. What you can control is whether you are giving the system the kind of creative variety it was built to reward, no matter which campaign type you run.

The New Rules for Creative Volume

For years, six ads per ad set was the common rule of thumb. That rule is gone. Many advertisers now run ten, twenty, or more ads in one ad set. But more ads is not automatically better, and this is where a lot of teams get the lesson half right (Jon Loomer, "Meta Andromeda and Creative Diversification").

The real difference to understand is concepts versus variants.

A variant is the same idea with small changes. Think of the same product photo with a different background, or the same headline with slightly different wording.

A concept is a genuinely different idea. A different visual style. A different type of customer being spoken to. A whole different format, like a customer testimonial video instead of a plain product shot.

Ten variants of one concept give the system almost nothing new to learn from. Ten ads spread across three or four real concepts, each with a couple of variants, give it a much stronger signal. Meta’s own ad tools support this too. Flexible Format ads let you add up to five headlines, five lines of text, and ten images or videos per ad. That’s useful for testing variants, but it does not replace having real concept variety.

Restructuring Campaigns and Ad Sets

A few habits follow directly from what we just covered.

Combine ad sets instead of splitting them too thin, and put real creative variety inside each one. Spreading your budget across too many narrow ad sets starves each one of the data both Andromeda and GEM need to personalize well.

Treat format as its own kind of variety, not just a delivery setting. Running the same idea as a video in one ad and as a static image in another gives the system a new kind of signal it did not have before.

Give new concepts real time before judging them. Both systems need real engagement data to learn how good a new ad actually is. Killing a new concept after 48 hours because it has not "proven itself" yet means judging it before the system had a chance to learn.

Do not confuse a full ad set with a varied ad set. Filling every open ad slot with small variants, just to look thorough, is exactly the pattern Andromeda was built to spot and ignore.

A Creative-Testing Routine Built for Andromeda and GEM

Here is a simple weekly routine that matches how these systems actually work.

1. Count concepts, not ads. Track how many real, different ideas are live in each ad set, not how many ad units are live. If every "new ad" this month is just a new variant of an old idea, your variety has not actually grown.

2. Keep concept testing separate from variant testing. Test new ideas against each other to find out what works. Test small changes within a winning idea to make it better. Mixing the two together makes it hard to learn anything useful from either one.

3. Give every concept a fair amount of time before you kill it. Long enough for the system to gather real data, short enough that you are not wasting money on something that is clearly not working. We go deeper on how to tell a slow starter from a real failure in our post on ad creative fatigue.

4. Refresh concepts, not just variants, on a schedule. Changing the background on the same old idea over and over does not give the system anything new, no matter how often you do it.

5. Review concepts every week, and review your whole account structure every month. Make this a regular habit, not a one-time fix after a rollout.

The Data: What Happens When an Algorithm Tests Your Creative for You

Wider’s Creative Loop is built around this same shift: it reads your ad account data and turns it into new creative to test, so genuinely new concepts keep entering rotation on a schedule instead of trickling in as an afterthought. Under Andromeda and GEM, that’s the difference that matters. Treat creative variety as infrastructure, not a one-time push, and you’re set up for how Meta’s systems actually work now.