Meta Introduces GenRecs Tokenization to Enhance Ad Recommendations

ME AI message,
Generative Recommenders (GenRecs) represent a paradigm shift from scoring each ad individually to simultaneously understanding ad content and user preferences to predict optimal outcomes for advertisers. Leveraging the inherent world knowledge of LLMs, ad matching becomes smarter, more precise, and continuously delivers compounded benefits for advertisers.
GenRecs Tokenization is the foundational data representation for achieving this goal. Instead of maintaining a dedicated ID for each item, GenRecs Tokenization represents each item using a set of tokens from a stable semantic vocabulary.
After integrating Meta’s Generative Ads Model (GEM) and Meta’s Adaptive Ranking Model, GenRecs Tokenization increased conversion rates for advertisers across various business scenarios by 0.35%.

GenRecs improves upon the current approach used by Meta’s ad recommendation system to make predictions. GenRecs stands for Generative Recommenders. It shifts the recommendation method from scoring each ad candidate individually to leveraging LLM architectures that simultaneously perform deep reasoning on ad content and user preferences. By combining pre-trained world knowledge with semantic understanding of the actual meaning of ads, the system no longer limits itself to matching click patterns, but instead uncovers the underlying patterns of user preferences. Even for entirely new products with no prior interaction history involving products, ads, or companies, the model can predict the most relevant ads for each user.
GenRecs Tokenization is the foundational data representation for achieving this goal. Language models do not maintain a unique ID for every concept in the world; instead, they represent any concept using a vocabulary composed of a small number of semantic tokens. Since tokens describe the meaning of items (such as ads, products, services, etc.), the model can establish connections between items that have never received interaction data and those it is familiar with. A newly launched "trail running shoe" can inherit relevant category knowledge the model has already learned, without having to construct a completely new representation from scratch with no signals.
GenRecs Tokenization introduces a structured vocabulary for ad recommendations, enabling GenRecs to cover all item types within the system and be used simultaneously for recall and ranking, including ads, products, homepage, and pixels.

Meta's advanced ad recommendation system continuously improves advertiser performance by learning about various items and their performance. LLMs can further enhance the recommendation system's performance by leveraging semantic tokenization across multiple ad scenarios:

Accelerate learning from new entities: semantic tokenization helps models infer relevance and establish connections by sharing concepts. New entities include newly created ads, segmented products, and homepages with limited interaction data.
Enable cross-campaign migration: Semantic tokenization establishes a stable vocabulary, allowing the model to transfer insights across advertisers' different campaigns.
Leverage the scaling advantages of LLMs: semantic tokenization allows these benefits to compound as performance improves along the LLM scaling curve.
Integrating world knowledge: semantic tokenization provides a direct interface for pre-trained language models to understand ad content and user intent.

After the project is migrated to semantic tokens, the model will learn from token patterns like an LLM. In production, this delivers the following effects:

Improved prediction accuracy in production: By integrating Meta’s GEM and Adaptive Ranking Model, GenRecs Tokenization increased conversion rate by 0.35%1.
Form a composite feedback loop: More accurate long-tail predictions enable better learning from long-tail ads, enriching the training data and further enhancing the model’s predictive capability for all ads, creating a continuous performance improvement feedback loop.
Introducing a new scaling approach: In production ranking and retrieval, tokens can scale along dimensions such as sequence length, training data, and model architecture.

1 Internal Meta data, measured in February 2026.
GenRecs Tokenization Introduction
The pre-trained multimodal model reads each ad, each product, and each homepage—including their text and images—and encodes these items into embeddings that carry world knowledge. The tokenizer combines these content signals with representations learned from user interaction behaviors, then quantizes the result into a shorter sequence of discrete tokens. Similar items share tokens, much like related words share roots. For example, there will be overlap between ads for running shoes from last season and this season, allowing new ad campaigns to immediately benefit from all the knowledge the model previously learned from last season’s products.
The core of GenRecs Tokenization is a three-stage processing workflow:

Encoding: Encode each item into the semantic embedding space.
Decompose: Decompose this embedding into mutually independent subspaces.
Quantization: Map each subspace to an entry in the codebook.

To maximize ad recommendation quality, different stage combinations should be employed based on the semantics of entities and features. For instance, for many ad representations, GenRecs Tokenization combines a learned encoder, residual decomposition, a learned vector codebook, and token parametrization using prefix n-grams.
Token quality can predict ranking accuracy.
Without considering downstream ranking benefits, tokenizer model quality can be evaluated using the following intrinsic metrics:

Cluster purity: Items sharing the same code often belong to the same category, enabling the model to receive clearer signals from each token;
Distribution balance: Projects are evenly distributed across different semantic IDs rather than being concentrated in just a few IDs.
Codebook utilization: A higher proportion of codes containing at least one item means the vocabulary can express more distinctions.

Among the 8 tokenizer versions tested, we found a strong correlation between these intrinsic quality metrics and downstream ranking prediction accuracy. Tokenizer versions that performed better on these metrics also achieved higher CTR prediction accuracy for the ad model.

Tokenization flywheel
Tokenization uncovers deep semantic patterns within vast product catalogs, maximizing the efficiency of high-traffic campaigns. For long-tail ads with limited exposure, tokenization enables ad recommendation models to generalize through shared content structures, restoring performance during cold start and enabling cross-entity transfer. When a sportswear brand launches its next season’s trail running shoes, tokenization allows the new product and its ads to build on knowledge the model learned from last season’s running shoes—even before anyone has seen the new ads.
By improving the prediction quality of long-tail ads, tokenization not only enables more accurate ranking of long-tail ads but also enhances training data, thereby improving prediction quality for all ads. This self-reinforcing flywheel improves model prediction quality with each iteration.

Recommended basics by LLM
Improvements to the tokenizer are not limited to the specific model targeted by its tuning. By optimizing codebook quality through diversity loss, dead code reset, and contrastive learning objectives, downstream performance can be measurably enhanced when generative recommendation models are trained using the generated tokens. Our work on LLM-native retrieval demonstrates the potential of this approach: a dual-tower architecture built on a shared LLM encoder achieves state-of-the-art recall quality using only a small fraction of training data, while maintaining efficient serving performance comparable to classical dual-tower models. Building on this, we are currently extending reasoning-aware ranking using the same vocabulary.
The Future of GenRecs Tokenization
We will continue advancing GenRecs Tokenization along three key directions to reach the optimal ROI boundary:
Enhance cold-start generalization: Long-tail ad inventory is the area where semantic tokens have the most significant impact and where they matter most to advertisers. We are expanding tokenization to cover all items, delivering better results for advertisers.
Prioritize tokenizer quality as a first-class goal: training aligned with interactive behavior, a unified cross-entity vocabulary, and codebook improvements are current key focus areas. Since intrinsic quality metrics can predict downstream ranking gains, we can continuously iterate on the tokenizer without waiting for full model training cycles.
Efficient service for LLM-native architecture: Fully verbalized tokenization provides the model with the richest possible perspective on ads, but the required context window is impractical at ad scale. We are combining compressed verbalized tokens with semantic IDs to maintain inference quality while keeping latency acceptable.
From a longer-term perspective, we are exploring whether the same vocabulary used today to improve ad ranking can be unified to represent content across different business scenarios. A shared token space covering text, image, audio, and video ad content would allow models to understand both user interests and advertiser offerings using a single representation, eliminating the need to concatenate separately trained embeddings.
As recommendation systems evolve from discriminative scoring to incorporate generative reasoning, their data foundation must transition from transient, one-time IDs to stable semantic representations. GenRecs Tokenization provides this foundation, transforming the complexity of Meta’s advertising ecosystem into a vocabulary that generative models can learn from, reason about, and use to generate content.
Acknowledgments
We thank Weifeng Cui, Ling Leng, Luke Simon, Hadar Ohana, Jimmy Chen, Minghai Chen, Chiyu Zhang, Dinesh Ramasamy, Debajyoti Roy, Shawn Shi, Gajanan Golegaonkar, Xiao Yang, Ashkan Aleali, Anuj Gargava, Xizhou Feng, Liquan Huang, Frank Shyu, Xiaoyu Chen, Kungang Li, Chonglin Sun, Matt Steiner, Sandeep Pandey, Santanu Kolay, Anuj Madan, Tak Yan, and all team members who contributed to the development and productionization of Meta’s first large-scale semantic tokenization project for advertising recommendations.
The article "GenRecs Tokenization: Reshaping Meta's Ad Recommendations with a Stable Model Vocabulary" was originally published on Engineering at Meta. (Source: MLion)