GenRecs Tokenization: Transforming Ads Recommendation at Meta with Stable Model Vocabulary

Why it matters

Meta has introduced GenRecs Tokenization, a data representation method that underpins its shift toward generative recommender systems in advertising. Rather than assigning a unique identifier to every ad, product, or page, the approach encodes each item as a short sequence of discrete semantic tokens drawn from a stable vocabulary. Because similar items share tokens, a newly launched product inherits what the model already knows about its category, which helps with items that have little or no engagement history. Meta reports that when combined with its Generative Ads Model (GEM) and Adaptive Ranking Model, the method produced a measurable lift in conversions across surfaces, according to internal data from February. The company frames the broader GenRecs effort as a move away from scoring individual ad candidates toward reasoning jointly about ad content and user preferences using LLM architecture. It also claims benefits including faster learning on new entities, transfer of insights across campaigns, and a feedback loop in which better tail predictions enrich training data. The technical pipeline involves encoding items into embeddings, decomposing them into subspaces, and quantizing via codebooks.