From the source
Meta is describing GenRecs Tokenization, a semantic vocabulary that lets its ads recommendation system reason about ad content instead of scoring each ad candidate on its own.
- The approach replaces a dedicated ID per item with tokens drawn from a stable vocabulary, so a newly launched product inherits what the model already learned about its category rather than starting from no signal.
- Paired with Meta's Generative Ads Model and Adaptive Ranking Model, Meta says GenRecs Tokenization delivered a +0.35% lift in conversions across surfaces, measured in February 2026 on internal data.
- The tokenizer reads each ad, product and page with pre-trained multimodal models, combines that content signal with engagement representations, and quantizes the result into a short sequence of discrete tokens through a three-stage pipeline of encoding, decomposition and codebook quantization.
- Meta says tokenizer quality can be judged on cluster purity, distribution balance and codebook utilization, and that across the eight tokenizer versions it tested, versions scoring better on those measures also predicted CTR more accurately.
- Meta says the gains compound: better tail predictions produce better lessons from tail ads, which enrich training data and improve predictions for all ads.
