Meta shared on its engineering blog how it has upgraded its main ad AI model, GEM (Generative Ads Recommendation Model). According to the company, over the past year GEM delivered +18.9% conversions and +12% CTR on ads across Meta's surfaces. Nothing changed in Ads Manager: the model works inside the ad delivery system.
What GEM is
GEM is a large "central" model that Meta first introduced in November 2025. It learns from signals across all of the company's ad platforms and jointly predicts clicks and conversions on Facebook, Instagram and other apps.
It's important to understand that GEM doesn't pick which ad to show by itself. It's too heavy for an auction that takes milliseconds. Instead, it passes its knowledge to lighter models, and those take part in ad selection.
Where the numbers come from and what's wrong with them
- +18.9% conversions and +12% CTR — across all Meta ad surfaces and objectives over the past year.
- +5.8% on Facebook, +4.1% on Instagram and +1.6% off-site conversions — this is uplift for a specific targeted group of users, not for everyone.
There's an important caveat: Meta didn't share the baseline, calculation method or sample size. These are the company's internal measurements and can't be independently verified. The numbers also can't be added together, and they certainly don't promise the same growth in your ad account.
What changed under the hood
The report covers three new developments:
- Kunlun — a new LLM-style model architecture. According to Meta, it delivers twice the quality gain on the same compute and lets the model account for a longer history of user actions.
- LoopFM — a mechanism that transfers GEM's knowledge and "memory" of user behavior to the lightweight models running in the auction. Meta says it doubles that transfer without extra latency.
- Solaris — a system that precomputes data for the user–ad pairs most likely to be needed in upcoming requests, so the heavy model doesn't slow down ad delivery.
The scale is impressive too: GEM is trained on several thousand of the latest GPUs, and training compute grew 4x over the past year.
One point some retellings got mixed up: right now GEM learns from ad signals. Accounting for organic behavior (for example, watching Reels without clicking ads) is something Meta describes as future work, not something already in place.
What it means for media buyers
- There's nothing to turn on. You can't select GEM in campaign settings; it already works for everyone.
- The algorithm relies more and more on signals. The more accurately the model predicts conversions, the more important it is to feed it quality data: a correct pixel, the Conversions API and the right optimization events.
- Don't change working funnels because of a press release. Creative, offer and landing page still matter more than any model update.
- Compare by money, not CTR. More clicks mean nothing on their own. Look at CPA, cost per confirmed lead, ROAS and profit over comparable periods.
Meta keeps scaling compute and training GEM on more and more signals. The direction is clear: fewer manual settings and more decisions made by the algorithm. For media buyers, that means the quality of conversion data is becoming one of the main levers for results.