Explainer

Meta's Andromeda and GEM, explained without the engineering

Andromeda is the part of Meta's ad system that makes the first cut: it narrows tens of millions of ads to the few thousand worth considering for one person. GEM is a very large model that works out what makes ads perform and teaches that to the smaller models that rank ads live. Meta says both reward a variety of creative, but in the sources we read it has not published a number of ads you need.

Last updated 9 October 2026

Two names, two different jobs

Meta does not choose an ad in one step. It first pulls a shortlist from everything that could be shown, then ranks that shortlist, then runs the auction. Andromeda and GEM sit at different points in that process, and most of the confusion comes from treating them as one thing.

What Meta has published about Andromeda and GEM
	Andromeda	GEM
What Meta calls it	A retrieval engine for ads.	The Generative Ads Recommendation Model, a foundation model for ads.
Its job	The first cut. Narrows tens of millions of candidate ads to a few thousand for one person.	Learns patterns from very large amounts of ad and non-ad activity, then passes that knowledge to the models that rank ads.
Does it pick your ad live?	Yes, at the shortlist stage.	No. Meta says it is too large to run at delivery time.
Where Meta describes it	Engineering blog, 2 December 2024.	Meta for Business, 27 March 2025; engineering blog, 10 November 2025.

Andromeda: the first cut

Meta's engineers introduced Andromeda on 2 December 2024. They describe retrieval as the first stage of ad delivery, where tens of millions of candidate ads are narrowed to a few thousand. Larger ranking models then score that shortlist. An ad that is not retrieved for a person is never ranked for them.

The system it replaced relied on separate model stages and many hand-written rules. Andromeda runs a much larger neural network on NVIDIA Grace Hopper chips and Meta's own MTIA hardware, and stores ads in a layered index so it can search them quickly. Meta reported a 6% improvement in recall, meaning how many of the right ads make the shortlist, and an 8% improvement in ads quality on selected segments.

Why build it? Meta's stated reason is volume. The same post says Advantage+ automation and generative creative tools increase the number of eligible ads, and that more than a million advertisers had used those tools to make over 15 million ads in a month. On the earnings call of 29 January 2025, chief financial officer Susan Li said Andromeda allowed a 10,000-fold increase in the complexity of the retrieval models, and that handling larger volumes of ads prepares Meta for advertisers creating and testing more of them.

It has kept changing. On the call of 29 October 2025, Li said Meta had combined retrieval and early-stage ranking into a single model, which raised ads quality on Facebook surfaces by 14%.

GEM: the teacher, not the picker

Meta named GEM in a Meta for Business post on 27 March 2025, saying an early version on Reels had lifted ad conversions by up to 5%. The fuller account came in an engineering post on 10 November 2025.

That post describes GEM as a foundation model built along the lines of a large language model and trained on thousands of GPUs. It learns from long histories of what people do with both ads and ordinary posts across Facebook, Instagram and business messaging. It is not designed to serve ads by itself. Its role is to act as a teacher: what it learns is transferred to the hundreds of smaller models that make live predictions.

Li made the same point on the October 2025 call: models such as GEM are not used at delivery time because their size would make that too costly, so Meta gets the benefit by passing their knowledge to lighter models. Meta credits GEM with a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in the second quarter of 2025.

What Meta says about how many creatives, and how varied

This is where published fact and common advice part ways. Meta has said three things.

Variety is the stated aim. The March 2025 post says businesses are uploading more creatives to support a diversification strategy, and that Andromeda lets Meta choose the right creative for each person. Meta's performance marketing page advises developing a variety of ad creatives and letting its systems match the message to the person.
Formats count as variety. Meta's ad creative page recommends a mixed-format strategy.
The system is built for more ads. Both the engineering post and the January 2025 call say Andromeda was designed to cope with a growing number of creatives.

None of the Meta sources listed below gives a number of creatives per ad set or campaign, a required refresh schedule, or a threshold for how different two ads must look. The performance figures Meta has published (6%, 8%, 14%, 5%, 3%) describe its own systems across the platform. They are not a promise of what an advertiser gains by adding ads.

So when you read that you need a fixed count of ads, that near-identical ads are merged above a stated similarity percentage, or that a given volume produces a given lift, treat it as someone's experience or inference. It may be sensible. It is not something Meta has stated in these sources.

What to do with this

A fair reading of Meta's own words is that different ads give the system more ways to match you to different people, and that small edits of one ad do less than a new idea. That second part is our inference, not a Meta rule. In practice:

Vary the idea, not the font. Change the hook, the angle, the customer you are speaking to, or the format.
Mix formats where you can: static images, video and carousels.
Run as many distinct ads as you can make well and afford to judge. A small budget spread over dozens of ads tells you little about any of them.
Judge by results in your own account. Test different ideas fairly and keep what works.
Be wary of any rule that comes with a precise number and no link to Meta.

Sources

Questions people ask

What is Meta Andromeda?

Andromeda is Meta's ads retrieval engine, introduced on its engineering blog on 2 December 2024. It handles the first stage of ad delivery, narrowing tens of millions of candidate ads to a few thousand for one person before larger models rank them. Meta reported a 6% improvement in recall and an 8% improvement in ads quality on selected segments.

What is Meta's GEM and does it choose which ad I see?

GEM is Meta's Generative Ads Recommendation Model, a very large foundation model trained on ad and non-ad activity across Facebook, Instagram and business messaging. It does not pick ads live: Meta says it is too large to run at delivery time, so it passes what it learns to the smaller models that rank ads. Meta credits it with a 5% rise in ad conversions on Instagram and 3% on Facebook Feed in the second quarter of 2025.

How many ad creatives does Meta say you need for Andromeda?

In the Meta sources we read, Meta does not give a number. It says businesses are uploading more creatives to support a diversification strategy, advises developing a variety of ad creatives and recommends a mixed-format strategy. Specific counts, similarity thresholds and promised lifts come from third parties, not from those Meta sources.

Try it on your own competitors.

$9 for 7 days. Paste a competitor's Page and see their best ad rebuilt for your brand.

Start 7-day trial · $9

Mockingbird is an independent product from Onwards Analytics and is not affiliated with, endorsed by or sponsored by Meta. Meta, Facebook, Instagram, Meta Ads Manager and the Meta Ad Library are trademarks of Meta Platforms, Inc. Meta changes its tools often; check Meta's own help pages for the current detail before you rely on it.