Meta Andromeda: What It Changes for Your Ad Creative

Meta Andromeda is the ads retrieval system Meta described in a December 2024 engineering post: the stage that narrows tens of millions of ads down to a few thousand candidates. Hayes Media, an eCommerce growth agency that runs Meta ad creative for DTC and Shopify brands, reads that as a reason to test more distinct concepts. That's our reasoning, and Meta doesn't state it.

This guide is for founders and marketing leads running Meta ads. Everything about Andromeda below comes from Meta's own post, listed under Sources at the end. We keep two things apart the whole way through: what Meta says, and what we conclude from it.

What is Meta Andromeda?

In Meta's words, "Meta Andromeda is a personalized ads retrieval engine". The post explains retrieval like this: "Retrieval is the first step in our multi-stage ads recommendation system." Its job is "selecting ads from tens of millions of ad candidates into a few thousand relevant ad candidates."

Then the next stage takes over: "In the following stage, larger and more sophisticated ranking models predict people and advertiser value to determine the final set of ads to be shown to the person."

So, as we read it, there's a shortlist and then a final pick. Retrieval decides which ads are in the running. Ranking decides which of those get shown. Your ad has to make the shortlist before ranking can consider it.

What does Meta say Andromeda changed?

Meta says it put far more personalization into the shortlist stage, and built that stage to handle many more ads. Four points from the post matter to anyone buying ads:

More personalization at retrieval. Before Andromeda, Meta says its retrieval systems "were only able to apply limited personalization". The new design enables "a meaningful increase of model capacity (10,000x) for enhanced personalization."

An index built for more ads. Meta says Andromeda "introduces an efficient hierarchical index to scale up to a large volume of ads creatives". It gives a reason too: with generative AI tools, "the number of ads creatives in Meta’s recommendation systems is expected to grow significantly."

Reported gains. Meta says deployment across Instagram and Facebook "has achieved +6% recall improvement to the retrieval system, delivering +8% ads quality improvement on selected segments." Those are Meta's own figures about its own system. They aren't an outside audit, and they aren't a forecast for your account.

A stated direction. Meta expected the model to move toward "a more diverse set of ad candidates", and wrote: "Increased ad diversity can improve people’s experience with ads and drive better advertiser outcomes." That's about the set of ads the system surfaces. It isn't an instruction to advertisers.

Keep the date in mind. The post is dated December 2, 2024, almost two years before we wrote this. It describes the system as Meta described it then. Meta may have changed things since, and this post can't tell you whether it has.

What doesn't Meta say about ad creative?

The post doesn't give any creative advice. It never mentions concepts, near-copies, or how many ads you should run. It doesn't say similar ads get penalized, and it doesn't set a target number of creatives.

It also leaves the inputs open. Meta refers to "precomputed ad embeddings and features", but it doesn't list what goes into them, so the post can't tell you how much your visuals or your script count. If someone tells you Andromeda rewards a specific number of ads or a specific format, ask where they got it. It isn't in this post.

Why does Hayes Media think this favors distinct concepts?

Everything in this section is our own reasoning from the mechanism. It isn't a Meta statement, and it isn't a measured result.

Meta says the system captures "Complex latent relationships between people’s interests, products, and services offered through ads". If that's what the shortlist runs on, we think what your ad is about is the part of it you control that matters most. That means who it speaks to, which problem it names, and why they should believe you.

Hayes Media's reasoning goes like this. Ten near-copies of one idea offer the system one argument, ten times. Five distinct concepts offer it five arguments, and each one may fit a different group of people. A retrieval system built to match ads to people has more to work with in the second case.

That's where our house line comes in: "Creative is the new targeting." It's our line, and Meta didn't write it. The Andromeda post fits with it, but it doesn't prove it.

We'd run distinct concepts even if this reasoning turned out to be wrong. A varied batch teaches you which arguments work, and it gives you more than one idea to fall back on when an ad wears out, which is the supply problem behind Meta ad creative fatigue.

What counts as a distinct concept?

A distinct concept makes a different argument to a different person. A near-copy makes the same argument with a new look. Here's the test we use: describe the ad in one sentence, leaving out the format. If two ads get the same sentence, they're one concept.

Here's an example for a made-up coffee brand:

Concept one. You're spending more on cafe coffee than you think, and this costs less per cup.

Concept two. The founder explains why she buys from a single farm.

Concept three. Someone who drinks instant coffee tries this and can't go back.

Near-copy of concept one. The same cost argument with a new color, a new font, or a different first line.

Near-copies still have a place. Once a concept wins, in our practice, varying the hook, the creator, and the setting is how you get more out of it. For more on what to put inside the frame, see what actually works in Meta ad creative. The change we're suggesting is about order: test the argument first, then the hook and format.

How should your creative testing change?

Test more ideas per batch, fewer variations per idea in the first round, and score both the ads and the concepts. Our full process is in the creative testing framework. These are the parts Andromeda's design makes us lean on harder:

Brief by concept. Write the argument and the audience at the top of each brief, before anyone thinks about format. Hayes Media sources and briefs creators for this, and the brand pays them.

Write the rules before launch. Hayes Media reports creative and media buying performance in new customer CAC or incremental ROAS against the CAC target the brand gives. Hayes never sets that target. A winner is an ad that meets it, and the spend at which you kill an ad gets agreed before launch.

Tag every ad at upload. Record the concept, hook, creator, format, and offer. Without tags you can have winners and no idea what they share.

Track hit rate at two levels. Hit rate is winners divided by ads tested. We also count concepts: how many of the concepts you tested produced at least one winner.

Read attention before conversions arrive. We look for a hook rate above 40% and check retention at fixed timestamps. Those are internal heuristics, not a platform standard. They tell you where an ad lost people, and they don't replace the CAC read.

Vary winners after they win. That's when near-copies earn their place, one change per version, so you can see which change did the work.

If you'd rather hand this over, it's what our Meta ad creative service does.

What does this look like in numbers?

These are made-up numbers, used only to show how to read a batch. They aren't a prediction or a client result. Say the brand's CAC target is $40, and two batches of 15 ads each run under the same rules.

Batch A. One concept with 15 variations. Three ads come in at or under $40 new customer CAC. Hit rate: 3 divided by 15, or 20%.

Batch B. Five concepts with three hooks each. Three ads come in at or under $40, two from concept two and one from concept four. Hit rate: 3 divided by 15, or 20%.

The hit rate is the same, but you know different things afterward. Batch A tells you one argument can work and which versions of it did best. Batch B tells you two arguments out of five can work, so 40% of your concepts produced a winner, and you have two separate ideas to build on. If concept two wears out, concept four is already there.

We made the hit rates equal on purpose, because we have no measured evidence that distinct concepts win more often. The point is that Batch B tells you more.

When this does not apply

Don't rebuild your account because of one engineering post. This advice won't help much in these cases:

Something outside the creative is broken. If your tracking, offer, or landing page has a fault, new concepts won't give you a clean read. Start with why Meta ROAS drops.

Your budget can't give each concept a fair read. Splitting too little spend across too many ideas leaves you with no answer on any of them. Test fewer concepts and agree the spend per concept first.

You're already running varied concepts and hitting your CAC target. Nothing in Meta's post asks you to change a program that works.

You want to know whether Andromeda caused a drop in your account. This post can't tell you that. It doesn't describe any effect on a specific account, so diagnose the account itself.

On fit, Hayes Media typically works with brands doing at least $100K per month in revenue. It's a guideline, not a gate: a smaller brand with a clear budget set aside to scale and real traction on Meta already qualifies.

What's the evidence behind our view?

It's experience, and we have no controlled test tying Andromeda to creative results. Our homepage says: "Our Performance Creative Process has produced winning ads that have spent $100,000's with on-target metrics." It also says "Trusted by 57+ eCom brands". Those claims, and our case studies, are self-published on our own site, and Hayes has no third-party review profile. Weigh them that way.

If you'd like us to look at your account, the call covers an audit of your recent creative: how many real concepts you're running, your hit rate against the CAC target you give us, and what to brief next. Book a discovery call. No onboarding fees. No lock-in contracts. No junior marketers.

Sources

Meta Engineering, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine", posted December 2, 2024. Fetched October 30, 2026. Every Andromeda quote above comes from this page.

Hayes Media homepage. Fetched October 30, 2026. Source of the two Hayes quotes above.