Meta's GenRecs Tokenization Delivers 0.35% Conversion Lift in Ads

• Meta's GenRecs Tokenization delivered a +0.35% lift in conversions across surfaces when paired with the Generative Ads Model (GEM) and Adaptive Ranking Model.
• The approach replaces dedicated item IDs with a stable semantic vocabulary, allowing the model to reason about ad content and user preferences simultaneously.
• Semantic tokenization enables new entities, such as newly created ads or niche products, to inherit learned knowledge from related items, improving cold-start performance.
• Intrinsic tokenizer quality metrics, including cluster purity and codebook utilization, strongly correlate with downstream ranking prediction accuracy.

Technical Implementation

• A three-stage pipeline encodes items into semantic embeddings, decomposes them into independent subspaces, and quantizes them into discrete tokens.
• Pre-trained multimodal models read ad text and images to create embeddings that carry world knowledge, which are then combined with engagement data.
• The system uses prefix n-grams combining learned encoders, residual decomposition, and learned vector codebooks for many ad representations.

Future Directions

• Meta is scaling tokenization coverage to improve performance for tail ad inventory where semantic tokens contribute most.
• The team is exploring a shared token space spanning text, image, audio, and video to unify representation across surfaces.