Meta's GEM Model: Why Creative Volume and Variety Now Win

You do not need to read Meta's engineering blog. I did, so here is the plain-English version and what I think it means for the ads you actually run.

On November 10, 2025, Meta's engineers described GEM, short for Generative Ads Recommendation Model. The post is dense, full of GPU counts and training tricks. Under the jargon is a simple idea with an operator consequence: the system that picks your ad is getting better at telling people apart. The more genuinely different things you give it to choose from, the more it has to work with.

What is GEM, in plain English?

Think of a head chef and a large kitchen. The head chef has tasted every dish ever served and learned which flavors each kind of diner tends to love. The line cooks each make one type of dish. GEM is the head chef. The hundreds of models that actually choose and rank ads are the line cooks. The chef teaches them, so they get better without tasting everything themselves.

Meta's own description: GEM is its most advanced ads foundation model, built with an approach inspired by large language models and trained across thousands of GPUs, and Meta calls it the largest foundation model for recommendation systems in the industry.¹

A few specifics from the post:¹

• GEM models user behavior sequences of up to thousands of events, rather than a short summary.
• It passes its learning to hundreds of user-facing vertical models, directly and through domain-specific foundation models.
• Meta says it is 4 times more efficient than its previous ads ranking models at driving ad performance for a given amount of data and compute.
• Since launch in Q2, GEM has contributed to ad conversions rising 5% on Instagram and 3% on Facebook Feed.

Notice what Meta did not publish: advertiser-level results. The conversion lifts are platform-wide, so they tell you where the platform is going, not what will happen in your account.

How does GEM fit with Andromeda?

Different layers of the same system.

Meta explained Andromeda in December 2024 as its retrieval engine (my take is in What the New Ad Retrieval Engine Means for Small Accounts). Retrieval makes the first cut, from a pool in the tens of millions down to thousands, and later ranking models pick the ads that show.² GEM sits behind the recommendation models, improving how they learn.¹

In the Q3 call, Meta's CFO said Meta had significantly improved Andromeda in Q3, with a 14% increase in ads quality on Facebook's services, according to a published transcript.³ I would not put a lot of weight on any single number. I would put weight on the direction: more modeling power at each stage, applied to the same job of matching ads and people.

Why would creative variety matter more?

Here is my reasoning. I want to be clear that this is inference, not something Meta has told advertisers.

Retrieval has to choose a few thousand ads from tens of millions for each moment.² Ranking has to choose the best. Both stages work by learning what kinds of ads go with what kinds of people. If you give the system five ads that make the same promise to the same type of person in different colors, you have given it one idea to match. If you give it five ads that make different promises, show different proof and open on different moments, you have given it five ideas to match to different people.

Meta's own Andromeda post points at this. It says the architecture is expected to move to an autoregressive loss function, which should yield more diverse ad candidates.² A system built to value diverse candidates gives you a reason to supply them.

One real idea beats ten recolors.

What counts as a different concept?

A concept is a different reason to buy, not a different look. Change at least two of these, and you probably have a new concept:

• The problem you open with. Time, cost, fear, status, frustration.
• The proof you lead with. A number, a story, a demonstration, an authority.
• The person you picture. The beginner, the veteran, the skeptic, the busy parent.
• The format. A talking head, a screen recording, a customer clip, a static before-you-buy checklist.

How do you build a concept library in practice?

The rule is concept first, then variation, then volume. A creative bank is where those concepts live. Here is the order.

What should you stop doing?

Stop three habits.

Stop recolor testing. If the only change is a button color or a font, you are using the budget to learn almost nothing. A real cadence is in Creative Is the New Targeting.

Stop starving ideas. Killing a new concept after a day because it has not beaten your best ad is a way of never finding the next best ad.

Stop shipping unreviewed volume. More creative is not an excuse for sloppier claims. Every variation still has to be true, which is a limit on how fast you can go. Meta says the number of advertisers using at least one of its video generation features was up 20% from the prior quarter,³ and those tools will make volume easy. Volume without review makes risk. (Guardrails for Meta's AI creative tools covers the review process.)

Ben Heath read Meta's Andromeda article in a video posted on November 11, 2025, and his headline advice leans toward volume:

He pairs that with more diversity of angles, desires and audiences, which is the part I would not skip. Where we differ is order. I want the concept settled first, then the variations, then the volume.

Plays here, from YouTube

How do you stay honest about the evidence?

Keep a creative log. For each concept, record the idea, the launch date, the spend and the result against your cost per purchase and cash collected. After a month, you will see which ideas carry weight, and you will stop trusting folklore, including mine.

Platform announcements are directional. They explain why a tactic might work, but your account is the experiment. The right response to GEM is not to change everything. It is to run a clean test of concept versus variation and let your own numbers decide, the way I set up testing in the systems I build for clients.

Keep building.Ekai