Meta has outlined three upgrades to the systems that decide which ad wins each impression: a larger GEM model, a reworked Andromeda retrieval system and GenRecs, a new generative recommendations model built on LLMs. Advertisers do not need to change any settings, but the update says a lot about where targeting and creative testing are heading.
What Meta changed
According to AdExchanger, the new GEM version is larger and reads twice as much user history as before. Andromeda, the retrieval layer that picks candidate ads, was overhauled with Nvidia to keep more ad data on GPUs and route ads to specialised expert models.
GenRecs is the new piece. Instead of treating each ad, product or page as a unique ID, it describes them with shared semantic tokens for attributes such as brand, colour, fit and style. A small generative retrieval model built on it predicts relevant ads from what a person engages with, and its output is added alongside the original retrieval model.
Why the cold-start angle matters
Meta's engineering blog says these tokens let the model relate a new or low-engagement item to similar items it already knows, such as a new product version learning from last season's model. Meta reported a 0.35% conversion lift from the tokenization work in February 2026 testing, measured alongside GEM and its Adaptive Ranking Model.
Meta's ads infrastructure VP Matt Steiner told AdExchanger that new ads still have to learn; the tokens shorten that phase rather than remove it.
The numbers Meta is quoting
Meta has credited its ranking work with 8.3% more ad clicks and a 15.7% lift in conversions on Facebook in Q2, and says GEM has delivered roughly 19% better conversions since last year. These are Meta's own platform-level figures and are not independently verified, so do not expect them account by account.
What to do about it
Steiner also said targeting matters less over time. In practice that means the creative and the product feed carry more of the targeting signal. Clear, distinct creatives and well-labelled catalog items give the semantic tokens more to work with, while narrow audience splits do less.
Expect the effect to show most on accounts that launch lots of new ads or products. If you test creatives in many small ad sets, consolidating into fewer, larger ones gives the ranking models more data per decision. Keep a stable baseline campaign running so you can tell a genuine change in performance from normal week-to-week noise, rather than reacting to every swing.
What it means for you
- No settings change needed; the upgrades apply automatically
- New ads and products may exit learning faster, but still need data
- Let creative and catalog attributes do the targeting work
- Treat Meta's lift figures as averages, not forecasts for your account
Sources
- AdExchanger: Meta has a new ad ranking system built on LLMs
- Engineering at Meta: GenRecs Tokenization
