Meta's engineering team has published an update on GEM, the Generative Ads Model that sits behind ad ranking on Facebook and Instagram. Meta says the 2026 version now works as one central model across its apps and objectives, and credits it with an 18.9% conversion improvement and a 12% CTR improvement. There is no setting to switch on: this is the delivery system every campaign already runs through.
What Meta published
In an October 13 post on its engineering blog, Meta describes GEM as a recommendation foundation model trained on data from all of its ads platforms. Earlier, Facebook, Instagram and Reels each leaned on their own optimisation systems, and each objective had its own prediction stack. The new generation jointly predicts clicks across all surfaces and conversions across major advertiser goals, from off-site conversions to app installs.
Meta's headline figures are an 18.9% conversion improvement and a 12% ad CTR improvement across its ad surfaces and objectives. Separately, it reports conversion uplifts over the past year for targeted users of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions.
How it reaches your campaigns
GEM itself is too large to score every ad in real time. Meta runs hundreds of smaller vertical models that serve ads, and GEM passes what it learns down to them. The post names two mechanisms: LoopFM, which compresses GEM's internal representations and combines them with a user's history, and Solaris, which precomputes user-ad embeddings for candidates likely to show up in future auctions.
Meta also says the 2026 model has 2.5x the complexity of earlier versions, doubled the production context length for user behaviour sequences, and doubled training efficiency over the past 12 months.
What to read into the numbers
These are Meta's own figures. The post gives no baseline, test design or time window for the 18.9% and 12% numbers, and no breakdown by country or vertical. Treat them as a direction of travel, not a lift you should expect to see in your own reporting.
The practical signal is clearer: Meta keeps moving optimisation into one central model that learns from behaviour across every surface. That rewards accounts that feed it clean conversion signals and enough creative variety, and it gives manual targeting and placement tweaks less to work with.
What to do now
If performance shifted in recent weeks without any change on your side, ranking updates like this are one plausible cause, so compare against a stable control campaign before rebuilding anything. Keep your Pixel and Conversions API events deduplicated and mapped to the right optimisation event, because better prediction only helps if the conversion data it learns from is accurate.
- Audit event match quality and deduplication before Q4 spend ramps up.
- Give broad campaigns several distinct creative concepts, not near-duplicates.
- Judge any change on a 7 to 14 day window, not daily swings.
What it means for you
- GEM runs behind every Meta campaign; there is nothing to opt into.
- The 18.9% and 12% figures are Meta's claims with no published method.
- Clean CAPI and Pixel signals matter more as ranking centralises.
- Creative variety gives the model more to work with than manual targeting.
Sources
- Engineering at Meta: Generative Ads Model (GEM), scaling frontier ads AI
- Common Thread Collective: Every Meta Ads Change in 2026
