Meta drops per-item IDs in ads ranking: a shared semantic vocabulary lifts conversions 0.35%

The ads recommender now describes each item with a set of shared semantic tokens instead of a dedicated ID for every ad, product, page and pixel, so the model can relate an item it has never seen engagement for to items it already knows. The prior setup kept a separate ID for every item, leaving a newly launched ad or a niche product with no signal and a representation built from scratch. Paired with Meta's Generative Ads Model and Adaptive Ranking Model, the tokenization delivered a 0.35% lift in conversions in production, measured on Meta internal data in February 2026.

A pre-trained multimodal model reads each ad, product and page, including text and images, and encodes it into an embedding that carries world knowledge; the tokenizer then splits that embedding into independent subspaces and maps each to a codebook entry, yielding a short sequence of discrete tokens. This three-stage pipeline, encoding into a semantic embedding, decomposing into independent subspaces and quantizing each subspace against a codebook, replaces assigning an ID per item, so similar items share tokens much as related words share roots. Across eight tokenizer versions tested, intrinsic measures such as cluster purity, distribution balance and codebook utilization correlated strongly with downstream ranking accuracy, and versions scoring better on those measures also predicted click-through rate more accurately.

The representation is in production across retrieval and ranking for ads, products, pages and pixels in Meta's ads recommendation system, paired with the Generative Ads Model and Adaptive Ranking Model. The 0.35% conversion lift comes from Meta internal data measured in February 2026. Stated next directions include scaling tokenization coverage across all items, treating tokenizer quality as a first-class objective, and combining compacted verbalized tokens with semantic IDs to keep latency acceptable.

The 0.35% conversion lift is Meta internal data, and no baseline or confidence interval is given. The link between intrinsic token quality and ranking accuracy comes from eight tokenizer versions and is a correlation, not a causal result. Fully verbalized tokenization needs context windows impractical at ads scale, and compacted alternatives are still being explored.