How Meta Improves Its Ad AI

Meta is expanding GEM, its main AI model for recommending advertisements. A recommendation system helps choose which ads people may find relevant. Earlier systems worked separately for different apps and advertising goals. The 2026 GEM model learns from information across Meta's advertising platforms. It supports goals from building awareness to encouraging people to buy.

Meta reports an 18.9% conversion improvement and a 12% ad click-through rate improvement. A conversion means a desired action, such as making a purchase. Click-through rate measures how often people click after seeing an ad. Meta also reports additional conversion gains across Facebook, Instagram, and off-site conversion. These figures describe reported improvements, not guaranteed results for every advertiser.

The new model design, Kunlun, helps GEM learn from longer activity histories. But running such a large model during every request would cost too much. It could also make people wait longer for an advertisement. So Meta passes useful learning to smaller models through a method called LoopFM. Solaris prepares some information before the smaller systems need it.

GEM combines learning across different places where Meta shows its advertisements. Examples include Facebook, Instagram, Reels, and Stories, which have different activity patterns. Information from tasks with many clicks can support tasks with fewer conversions. Shared learning works together with systems that understand each platform's particular patterns. This approach replaces the earlier separation of platforms and prediction goals.

Training a large recommendation model also requires changes to hardware and software. Meta reports that training efficiency doubled to 20-25% model FLOPS utilization. Total training computation increased by 4x during this process. The company changed several parts of the training system together. Its next plans include longer activity histories and more platform information. The main goal is better ad matching without adding delays to live requests.