Meta GEM Model: How Facebook Ad Delivery Works in 2026

In October Meta updated GEM, the model that decides who sees each ad on Facebook and Instagram. For a media buyer this is not abstract news from an engineering blog. It shapes how a campaign should be built now, which manual settings still make sense, and why the creative matters more than the targeting.

What Is Meta's GEM Model?

GEM (Generative Ads Recommendation Model) is Meta's main ads AI model, built on the same principles as large language models. Previously each placement (Facebook Feed, Instagram, Reels, Stories) and each objective had its own model. Now one model learns from a person's behaviour across all of Meta's apps at once: what they watch on Instagram helps predict what they will do on Facebook, and the other way round.

How Is GEM Different from Andromeda?

Andromeda and GEM work at different stages of ad delivery. Andromeda picks, out of millions of ads, the ones that could suit a person at all, and it is Andromeda that reads the visual pattern of a creative. GEM then ranks that shortlist and decides what is shown first. Put simply, Andromeda decides whether you get into the game, and GEM decides whether you win it. We covered the controls Andromeda brought in a separate piece: Meta Andromeda in 2026.

What Meta Says About the Results

According to Meta itself, at launch in 2025 GEM delivered a 5% increase in conversions on Instagram and 3% on Facebook Feed. In its October 2026 update the company reports an 18.9% conversion improvement and a 12% CTR improvement across its surfaces and objectives. These are Meta's own figures, not an independent measurement. But the direction is clear: the platform is investing in the model and expects advertisers not to get in its way.

How to Build a Campaign for GEM

Broad targeting. Broad or Advantage+ Audience, no narrow interests. The model sees a long user history and finds the people interested in the ad by itself.

A compact structure. Fewer campaigns and ad sets, more budget in each, no unnecessary duplicates. The model learns from signals, and a scattered budget gives it less data at every point.

A clean conversion signal. The pixel together with the Conversions API, with event value passed through. Optimise for an event that reaches roughly 50 a week.

Leave the campaign alone. Give it 2-3 days in the learning phase and raise the budget by 20-30% at a time. We scale the same way: +20-30% once a day, at the start of the day in the ad account's time zone.

Why Is the Creative Now the Targeting?

GEM matches ads to people based on what the ad itself contains. So the audience is defined less by ad set settings and more by what you show and how. Creatives that differ in meaning (a different pain point, a different offer, a different angle) open different audiences to the model. Ten copies with a different button colour open nothing. Keep formats varied too: vertical video for Reels, static images, carousels.

Should You Still Exclude Placements?

You can, but there is less manual control every month. Some ad accounts already show a warning that excluding placements, platforms and devices at ad set level will soon no longer be possible. On top of that, a checkbox that is on by default lets Meta spend up to 5% of the budget on each excluded placement.

What is left:

• Placement controls in the ad account settings. They apply to all campaigns, and changes to campaigns already running can take up to 48 hours.
• Value rules at ad set level: the bid on an unwanted placement can be lowered by up to 90%.
• After every ad set duplication, check the 5% checkbox again, because copying sometimes resets settings to the default.

Launching a Campaign Through Meta AI

A text-based launch has appeared in beta: you describe the goal, budget and timing, and Meta AI proposes the structure and settings itself. Buyers who tested it say that with a specific prompt it takes about 10 minutes. But one of them had the account hit a checkpoint twice, and the second one was a selfie check. For now the tool is worth testing on spare accounts, not on working ones.

What This Means for a Gambling Media Buyer

GEM rewards exactly what is hardest to get in gambling: a stable signal and time to learn. Ad accounts often live shorter than the model needs to learn, so everything that outlives a single account gains value: a creative library with genuinely different angles, a clean conversion signal and a compact structure that is easy to move to a new account. Our takeaway: ABO stays for finding creatives, but winners should be consolidated into CBO earlier and the budget kept in one place. This is exactly how we teach campaign building in our Facebook media buying program.

If you want to build campaigns for Meta's new models on live launches rather than by trial and error, we will show you how the process works.