TL;DR: GEM is the large model at the centre of Meta's ad ranking. It reads a person's history and an ad's content, and it learns from Facebook, Instagram and Reels at once. It passes what it learns to the smaller models that pick the ads people see. Meta's October 2026 figures: 18.9% more conversions and a 12% higher click-through rate across its surfaces. No setting turns it off. What you control is the material it learns from. That means ads that say what the product is, conversion events it can read, and ad sets left in place long enough to learn.
What is Meta GEM?
GEM stands for Generative Ads Recommendation Model. Meta first named it in a business post of March 27, 2025, and described it in full in an engineering post on November 10, 2025. Meta calls it its "most advanced ads foundation model", "built on an LLM-inspired paradigm and trained across thousands of GPUs". Meta also calls it the largest foundation model for recommendation systems in the industry.
A foundation model is one large model trained on everything, whose knowledge is then handed down to smaller models. Meta says GEM teaches "hundreds of user-facing vertical models", the ones that score ads in the live auction. GEM itself is too large to run inside an ad request. The October 2026 post says so in plain words: evaluating a model of this size "inside a live ad request is not affordable". That is why GEM works in the background. Its predictions reach the auction through the smaller models and through results it computes ahead of time.
What GEM learns from
Meta's November 2025 post says GEM learns from "ad content and user engagement data from both ads and organic interactions". Its inputs fall into two groups.
- Sequences: a person's activity history, which Meta says can run to "thousands of events" of ad and organic interactions.
- Attributes: Meta's examples are "age, location, ad format, and creative representation", so the ad's own content is among them.
The model learns across surfaces. Meta's example is that engagement with video ads on Instagram improves predictions for ads in Facebook Feed. The 2026 version doubled the length of engagement history it reads in production.
What changed in the 2026 version
Meta's engineering post of October 13, 2026 describes what it calls the 2026 generation of GEM. The change in one sentence, Meta's: GEM "shifts from isolated per-domain optimizations, such as Facebook, Instagram, Reels, to a single central foundation model". Until then each surface and each goal had its own prediction system. Now one model "jointly predicts clicks and conversions across all surfaces" and across every advertiser goal, "from off-site conversions to mobile app installs".
| First version (2025) | 2026 version | |
|---|---|---|
| Announced | March 27, 2025; engineering post November 10, 2025 | Engineering post October 13, 2026 |
| Scope | One model, learning across surfaces, teaching surface-specific models | One central model predicting clicks and conversions for every surface and goal |
| Meta's figures | +5% ad conversions on Instagram, +3% on Facebook Feed (second quarter of 2025) | +18.9% conversions and +12% click-through rate across surfaces and goals. A further +5.8% (Facebook), +4.1% (Instagram) and +1.6% (off-site conversions) for targeted users over the past year |
| Model size | 4 times more efficient than Meta's earlier ranking models for the same data and compute | 2.5 times the model complexity of the earlier GEM, on a new architecture Meta calls Kunlun |
Two technical lines from the October post matter to advertisers. First, "the generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks". In plain words, what the model learns from clicks, which are plentiful, improves its predictions of sales, which are rare. Second, a system Meta calls Solaris computes GEM's prediction for a person and an ad shortly before the auction needs it. The large model's view reaches the live request without slowing it down.
GEM, Andromeda and GenRecs: who does what
Meta's delivery runs in stages, and three names come up in its engineering posts.
Andromeda is the retrieval stage. It narrows tens of millions of eligible ads to a few thousand candidates. The Andromeda post covers it.
GEM is the central model behind the ranking stage. The ranking models that score those few thousand candidates learn from it. Meta names its Adaptive Ranking Model, described on March 31, 2026, next to GEM in that stage.
GenRecs Tokenization, described on October 19, 2026, changes how these models see every item. Instead of an ID with no meaning, each ad, product and page gets tokens that carry meaning. Models that "read each ad, product, and page, including its text and images" build those tokens. Meta's example: a newly launched trail shoe "inherits what the model has already learned about that category, rather than starting a new representation with no signal". Meta measured a 0.35% lift in conversions from it in February 2026, paired with GEM and the Adaptive Ranking Model.
What it means for your account
Five consequences follow from Meta's descriptions. They are my reading of the posts, not Meta's instructions.
The model reads your ad's content. Creative representation is one of GEM's inputs, and GenRecs describes each ad by its text and images. An ad that says what the product is gives the model something to match. A new ad for a product line Meta already knows starts from what the earlier ads taught it.
One model covers every placement. The 2026 GEM learns Facebook, Instagram and Reels together, and clicks on one surface improve conversion predictions on another. Splitting placements into separate ad sets, or excluding surfaces to "protect" results, removes signal the model would have used. Advantage+ placements let the cross-surface learning work.
You do not need to optimise for clicks to feed the system. Meta says click data already transfers into conversion predictions. Optimise for the result you want, even a rare one. A sales campaign with few sales still benefits from the clicks the model sees elsewhere.
Your conversion events are training data. A person's purchases sit in the history GEM reads. If your pixel or Conversions API misses events, the model learns from fewer of your outcomes. The Meta Ads Data Advisor guide shows how to check that your events arrive.
Stability pays. GEM reads thousands of events of history and computes predictions ahead of the auction. An ad set that stays in place accumulates the history those predictions rest on; a significant edit sends it back through the learning phase.
Do:
- Write ads whose text and image say what the product is and who it serves.
- Keep placements automatic and judge results by cost per result over weeks.
- Check your conversion events before you blame delivery.
Do not:
- Create one ad set per placement to compare them.
- Switch a sales campaign to optimise for clicks to "get data".
- Rebuild ad sets after one bad week.
The ad sheet in Adsap launches a set of distinct creatives into one ad set in one pass. The performance dashboard shows cost per result by ad over any period.
Frequently asked questions
Generative Ads Recommendation Model. Meta's engineering posts also shorten it to Generative Ads Model. It is a foundation model, so it teaches the smaller models that score ads in the auction rather than scoring them itself.
No. Andromeda is the retrieval stage that narrows tens of millions of ads to a few thousand candidates. GEM is the central model behind the ranking stage that scores those candidates. Meta's posts list them as separate parts of the same delivery chain.
No. Advantage+ is the name of Meta's automation products in Ads Manager, such as Advantage+ placements and Advantage+ creative. GEM is a model inside Meta's delivery system. It runs for every advertiser whether or not they use Advantage+ products.
No. There is no setting for it in Ads Manager and no field for it in the Marketing API. It is part of Meta's ad delivery for every advertiser on Facebook and Instagram.
No. Audiences, placements and budgets work as before. What changed is the quality of the predictions made inside those settings, and the fact that one model now learns from every surface at once.
Sources
All Meta pages read on 2 November 2026.
- Engineering at Meta, Meta's Generative Ads Model (GEM): The Central Brain Accelerating Ads Recommendation AI Innovation, November 10, 2025. The definition, the inputs, the "hundreds of user-facing vertical models", the Instagram-to-Facebook example, the 5% and 3% conversion figures and the 4 times efficiency figure.
- Engineering at Meta, Meta's Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI, October 13, 2026. The single central model, the 18.9% and 12% figures, the per-surface uplifts and the 2.5 times complexity. Also the click-to-conversion transfer sentence, the doubled history length, Solaris, and the line on live requests.
- Engineering at Meta, GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary, October 19, 2026. The token description of items, the trail shoe example and the 0.35% figure (Meta internal data, February 2026).
- Engineering at Meta, Meta Adaptive Ranking Model, March 31, 2026, as referenced in the GenRecs post.
- Meta for Business, AI Innovations in Meta's Ad Ranking Driving Advertiser Performance, March 27, 2025: the first public mention of GEM.
- Adsap: Meta Andromeda, explained, Meta Ads Data Advisor, ad sheets, the performance dashboard and the learning phase glossary entry.