Meta says its Generative Ads Model, known as GEM, has delivered an 18.9% conversion improvement and a 12% click-through rate improvement across its ad surfaces and objectives. Meta’s engineering team reported the figures on October 13.
The post describes the 2026 generation of GEM as a single central foundation model for ads recommendation. Earlier versions were tuned for individual surfaces such as Facebook, Instagram and Reels. According to Meta, the new version learns from data and signals across all of its ads platforms and jointly predicts clicks and conversions across surfaces and advertiser objectives, from off-site conversions to mobile app installs.
Meta also reports conversion uplift of 5.8% on Facebook, 4.1% on Instagram and 1.6% on off-site conversion for targeted users over the past year. On the engineering side, it cites a new architecture called Kunlun, 2.5x model complexity growth from earlier versions, and a post-training technique called LoopFM that transfers GEM’s representations to the models that serve ads.
Why it matters
These are Meta’s own figures from an engineering write-up, given as results across its ad surfaces and objectives. They do not predict the result on an individual campaign, so an advertiser cannot expect 18.9% from this post alone.
In our reading, the useful part for media buyers is the structure. One central model across surfaces means a signal that improves Instagram ranking can also influence Facebook, so account-level changes in creative or conversion data may travel further than a single placement report suggests. A holdout test run before and after your own conversion setup changes is the way to find out.
The post sits beside Meta’s recent product work on automated ad tools, including the assistant changes we covered in Meta’s Ad Assistant Gains Memory While Agentic Actions Stay in Testing. Related coverage: Amazon, Meta, Discord and Google Hand More Ad Buying to Software.
Source: Engineering at Meta, Meta’s Generative Ads Model (GEM)
