GEM is the model now teaching every ad-ranking system on Facebook and Instagram. Here is how it works, what Meta says it has delivered, and the account changes that help it work for you.
Switch between eras, then tap a surface to see what Meta reports for it.
GEM (Generative Ads Recommendation Model) is Meta’s largest ads foundation model, trained at large-language-model scale on ad and organic engagement data. It does not serve ads directly. It teaches the hundreds of smaller ranking models that do. Meta reports it has lifted conversions by 18.9% and click-through rate by 12% across its ad surfaces. For advertisers, it rewards clean conversion data, varied creative, broad placements and consolidated campaigns.
GEM is a foundation model for ads recommendation. Meta built it on an approach borrowed from large language models: train one very large model on as much data as possible, then pass its knowledge down to smaller, faster models. Meta calls GEM its “central brain” for ads.
The reason for the split is speed. Ranking an ad happens in milliseconds, and a model of GEM’s size is too expensive to run on every request. So GEM trains offline, and the user-facing models, which Meta calls vertical models, learn from it. When people say “the GEM algorithm”, they mean this whole system: GEM plus the ranking models it improves.
GEM learns from ad content and engagement data from both ads and organic interactions. Meta sorts that data into two kinds of features and handles each with its own attention mechanism, while letting them learn from each other.
Older systems compressed that history into one small summary vector before using it, which Meta says risks losing important engagement signals. GEM’s 2025 design, InterFormer, kept the full sequence and alternated between sequence learning and cross-feature layers. The 2026 version replaces it with a new architecture called Kunlun.
Drag to lengthen a person’s engagement history. Older models squeeze it into a fixed summary. GEM keeps the sequence.
Right: Meta reports that prediction quality follows a power law with sequence length. Curve shape is illustrative, not Meta data.
Facebook Feed, Instagram, Reels and Business Messaging each have different user behaviour. Older systems either treated each surface separately and missed cross-platform patterns, or treated them all the same and ignored the differences. GEM learns from interactions across surfaces while keeping predictions tailored to each. Meta’s example: engagement with Instagram video ads can improve predictions for Facebook Feed ads.
In 2026 Meta went further. Each surface and objective used to have its own prediction system. The new GEM is a single central model that jointly predicts clicks across all surfaces and conversions across the major optimisation goals. A shared expert learns patterns common to every domain, and domain-specific experts handle what is unique to each. Meta says learnings from data-rich click tasks now transfer into data-sparse conversion tasks.
GEM only matters if its knowledge reaches the vertical models. Meta uses two routes:
The techniques behind this are knowledge distillation (the small model learns to match GEM’s predictions), representation learning (GEM produces compact features the small models can use) and parameter sharing (small models reuse parts of the large one). A lightweight component called the Student Adapter corrects GEM’s outputs with the latest real outcomes, so the vertical models are not learning from stale predictions. Meta reported this combination is twice as effective as standard knowledge distillation.
The 2026 update added two more methods. LoopFM passes GEM’s internal representations and its memory of a user’s history to downstream models, which Meta says doubles intelligence transfer at no extra compute cost or latency. Solaris precomputes GEM’s view of likely user-ad pairs just before a request needs them, so ranking models get near-real-time knowledge from the large model.
Meta has published results twice. The figures below are Meta’s own, averaged across its platform. Individual accounts will vary.
Compare what Meta reported at launch and one year later.
The training numbers matter to advertisers for one reason: Meta says that in Q3 2025, architecture changes doubled the performance gained from a given amount of data and compute. In plain terms, every extra GPU Meta puts into GEM now buys more ad performance than before, so Meta has a business reason to keep scaling it.
Advertisers often mix the two up. They work at different stages of ad delivery.
| Andromeda | GEM | |
|---|---|---|
| Stage | Retrieval | Ranking |
| Job | Narrows the pool of eligible ads to a candidate set for each person | Improves the models that score and order those candidates |
| Introduced | December 2024, alongside Advantage+ | Launched 2025, detailed November 2025 |
| What it means for you | More, more varied creative gives it more to retrieve | Better signals give it better predictions |
Both point the same way. The system needs varied creative to choose from and reliable conversion data to learn from. For more on the creative side, see our guide to why Meta Ads ROAS drops and how to fix it.
Meta does not publish advertiser guidance in its engineering posts. The steps below are our reading of how the model learns, applied to the accounts we run.
Meta names sparse signals as one of GEM’s core challenges: billions of interactions a day, but few clicks and conversions. Every conversion you fail to report is missing training data. Run the Conversions API alongside the pixel and pass purchase value.
GEM treats a representation of the creative as a feature, and Meta plans to extend it across text, images, audio and video. Different concepts, not small variations of one ad, give both Andromeda and GEM more to match against different people.
The 2026 model learns across surfaces in one system. Restricting an ad to one placement limits where it can find the people the model predicts will convert.
GEM learns from long histories on the user side. Your side still needs enough conversion volume per ad set for delivery to stabilise. Fewer, better-funded ad sets usually beat many thin ones.
GEM trains on organic interactions as well as ads. How people engage with your organic posts and Reels is part of the picture Meta builds.
GEM optimises jointly across awareness, engagement and conversion. Judge it on blended ROAS, contribution margin and new-customer cost, not one campaign’s in-platform ROAS.
Enter your numbers. This applies Meta’s platform-wide average to your account as a rough illustration, not a forecast.
Uplift figures: Meta, Oct 2026, conversion uplift for targeted users over the past year. Same spend assumed.
Meta says it will extend GEM to learn from all its platform data, organic and ads, across text, images, audio and video, and to reason over longer engagement histories. Its 2025 post also mentions inference-time reasoning and “agentic, insight-driven advertiser automation”. Expect more of campaign setup to move into automated products like Advantage+, with the advertiser’s job shifting further toward creative, offers and data quality.
GEM, the Generative Ads Recommendation Model, is Meta’s largest foundation model for ads recommendation. It is trained at large-language-model scale on ad and organic engagement data and transfers what it learns to the hundreds of smaller models that rank ads on Facebook, Instagram and other Meta surfaces.
No. GEM runs inside Meta’s ad delivery system and has no setting. Advertisers influence it indirectly through the signals they send: conversion data, creative, placements and campaign structure.
In November 2025 Meta reported a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025. In October 2026 it reported an 18.9% conversion improvement and a 12% click-through-rate improvement across its ad surfaces and objectives, plus conversion uplift for targeted users of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions over the past year.
Andromeda is Meta’s ads retrieval engine, which narrows millions of possible ads to a candidate set for each person. GEM is a ranking foundation model whose knowledge improves the models that score and order those candidates. Retrieval decides which ads are considered, ranking decides which one is shown.
Yes. Meta says GEM is trained on ad content and user engagement from both ads and organic interactions, and plans to extend it to learn from organic and ad content across text, images, audio and video.
Send complete conversion data through the Conversions API, run a wide variety of creative concepts and formats, allow broad placements so cross-surface learning can work, avoid splitting budget into many small ad sets, and judge results on blended business metrics.
The 2026 generation moved from per-surface optimisation to a single central model that jointly predicts clicks and conversions across surfaces and objectives. It uses a new architecture called Kunlun with 2.5x the model complexity, doubled the engagement history length in production, and added post-training methods called LoopFM and Solaris to pass fresher knowledge to the serving models.
Related reading: Retargeting in 2026 · UGC ads and nano influencers · D2C ads agencies in Mumbai
We audit signal quality, creative volume and structure, then build the plan with you.