Meta GEM
- Also called GEM, Generative Ads Recommendation Model, Generative Ads Model
- Updated October 31, 2026
Meta GEM (Generative Ads Recommendation Model) is the large foundation model behind Meta's ad recommendations. It learns from ad engagement, clicks and conversions across Facebook, Instagram and Meta's other surfaces, then passes what it learns to the smaller models that choose ads in real time. It recommends ads; it doesn't make them.
How does Meta describe GEM?
Meta's 2025 engineering post calls it "Meta's most advanced ads foundation model, built on an LLM-inspired paradigm and trained across thousands of GPUs". The October 2026 update calls it "the industry's largest recommendation system (RecSys) foundation model" and describes what changed:
The 2026 generation of GEM shifts from isolated per-domain optimizations — such as Facebook, Instagram, Reels — to a single central foundation model.
That one model "jointly predicts clicks and conversions across all surfaces, including Facebook and Instagram, and across every key advertiser objective from off-site conversions to mobile app installs." Meta adds that "Each surface and optimization objective previously had its own independent prediction system."
Does GEM choose the ads people see?
Not directly. The 2026 post says "Serving a model at this scale in real time is prohibitively expensive and drives up latency", so GEM's value "depends on how effectively its intelligence transfers to the hundreds of vertical models that serve ads in real time." Meta's overview for advertisers puts it in one line: "GEM learns from trillions of signals and shares what it learns with the smaller models that choose ads in real time."
How is GEM related to Andromeda?
Meta describes them as separate parts of one ads system. Its October 2026 summary calls Andromeda "our ads retrieval engine", the step that narrows "a pool of tens of millions of ads down to the shortlist we consider showing someone." GEM, in the same summary, passes "what it learns centrally and its historical memory to the models that rank ads in real time, across every surface and every objective." Meta's engineering posts on the two don't mention each other.
What results does Meta attribute to GEM?
| Where | Meta's words |
|---|---|
| 2025 engineering post | "a 5% increase in ad conversions on Instagram and a 3% increase in ad conversions on Facebook Feed in Q2" |
| 2026 engineering post | "18.9% conversion improvement and 12% ad click-through rate (CTR) improvement across Meta's ad surfaces and objectives" |
| 2026 engineering post | "5.8%, 4.1%, and 1.6% conversion uplift on Facebook, Instagram, and off-site conversion for targeted users over the past year" |
| 2026 advertiser summary | "Since 2025, GEM has delivered up to a 18.9% conversion improvement" |
The per-app figures are "for targeted users", and the advertiser summary words the headline figure as "up to". All four are Meta's measurements.
What it means for your ads
- There is no GEM setting. Meta says the gains from these models "reach your account with no work from you."
- Your creative is one of its inputs. The 2025 post lists "ad format, and creative representation" among the ad attributes GEM learns from, next to activity history.
- What people do on one placement informs predictions on another. Meta's 2025 example: GEM can "use insights from Instagram video ad engagement to improve Facebook Feed ad predictions". The 2026 version learns "from data and signals across all of Meta's ads platforms".
- Meta names one group the 2026 version helps: "Campaigns with few conversions now benefit from what GEM learns about other clicks data, so advertisers with deep-funnel objectives gain better performance."
Common mistakes
- Reading "generative" as ad generation. Meta describes GEM as a model that improves performance "by enhancing other ads recommendation models' ability to serve relevant ads". Tools that generate images or text for an ad are Advantage+ creative features.
- Treating 18.9% as a forecast for your account. It is a system-wide figure, and Meta's advertiser summary says "up to".
Sources
- Meta's Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI — Engineering at Meta, October 13, 2026 (accessed October 31, 2026)
- Meta's Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation — Engineering at Meta, November 10, 2025 (accessed October 31, 2026)
- Inside the Three New AI innovations in Ads Ranking Improving your Ad Performance — Meta for Business, October 28, 2026 (accessed October 31, 2026)
- How Meta's Ads System Matches Your Ad to the Right Person at the Right Time — Meta for Business, October 28, 2026 (accessed October 31, 2026)
Meta restates GEM's results with each update, so check the source before quoting a figure.
Terrads makes each placement you pick at its own size, composed for its frame, so every surface GEM learns from gets an ad built for it.