Meta replaces ad IDs with a semantic token vocabulary to sharpen recommendations

Meta’s engineering team detailed GenRecs Tokenization, the data layer behind what it calls Generative Recommenders. Instead of giving every ad, product, page or pixel its own dedicated ID, the system describes each item with a short sequence of tokens drawn from a stable semantic vocabulary. Similar items share tokens, so a new ad can inherit what the model already knows about related items. The tokenizer works in three stages: encode an item into a semantic embedding using pretrained multimodal models that read its text and images blended with engagement signals, decompose it into independent subspaces, then map each subspace to a codebook entry.

Paired with Meta’s Generative Ads Model (GEM) (GEM) and its Adaptive Ranking Model, Meta says the approach delivered a 0.35% lift in conversions across surfaces, based on internal data measured in February 2026. That sounds modest, but at Meta’s scale a fraction of a percent in conversions represents a large amount of advertiser spend working harder.

Item IDs are brittle: a new product starts with nothing, and a niche ad with little traffic never gathers enough signal to be ranked well. With semantic tokens, a new trail shoe looks like last season’s trail shoe from the first impression, a real fix for cold start that most benefits smaller advertisers with thin data. Meta also reports that across eight tokenizer versions, intrinsic measures such as cluster purity, balanced distribution and codebook use tracked downstream ad-click prediction accuracy, letting the team improve the tokenizer without waiting for full model training cycles.

The shift reflects a broader move in recommendation systems from scoring candidates on click history toward language-model architectures that reason about what an item is and what a person wants. Meta says it is building LLM-native retrieval on the same vocabulary and reasoning-aware ranking on top, combining compact verbalized tokens with semantic IDs to keep latency acceptable.

This is a system for understanding people better in order to sell to them. Meta describes a feedback loop where better predictions on long-tail ads improve training data, improving predictions for every ad; such flywheels are efficient and concentrate advantage, making it harder for others to compete for ad dollars. Meta’s roadmap of a single token space across text, images, audio and video would let models reason about a person’s interests and an advertiser’s offer in one representation, a deeper preference profile than click histories, and the post says nothing about how users can see or shape it. Meta’s ad systems have also faced scrutiny over harmful ads that slipped through review (San Francisco order). More in our Meta coverage, marketing coverage and research coverage.