Meta’s engineering blog described GenRecs Tokenization on October 19 and reports that, paired with its Generative Ads Model (GEM) and Adaptive Ranking Model, it has delivered a 0.35% lift in conversions across surfaces for advertisers.
What happened
Instead of keeping a dedicated ID for each ad, product or page, the system represents each item with a short sequence of tokens from a shared semantic vocabulary. A three-stage pipeline encodes the item into an embedding, splits that into subspaces and maps each one to an entry in a codebook, so similar items share tokens. Meta says a newly launched trail shoe inherits what the model learned from last season’s shoe instead of starting with no signal. The post names new ads, niche products and low-engagement pages as the cases that gain. A footnote says the 0.35% figure is Meta internal data, measured in February 2026.
Why it matters
The change targets cold start, where an ad or product has too little engagement history to rank well. Meta places the gain in tail inventory, which matters to advertisers that launch new creative or catalog items often. Across eight tokenizer versions, it reports that higher scores on cluster purity and codebook utilization went with better click-through-rate prediction.
One detail worth reading closely
In our reading, 0.35% is a small number next to the 18.9% conversion lift Meta reported for GEM, which we covered in Meta Reports GEM Ads Model Lifted Conversions 18.9% Across Its Ad Surfaces. The two figures measure different things: the new one is the increment from a data representation layered on top of GEM and the Adaptive Ranking Model. Meta gives no method beyond the internal-data footnote, and the post makes no claim about what any one campaign will see. Meta lists a shared token space across media types as a direction. For the assistant side of the stack, see Meta’s Ad Assistant Gains Memory While Agentic Actions Stay in Testing.
Source: Engineering at Meta
