Meituan Search Advertising: Evolution of Recall Strategies and Generative Approaches

This article summarizes Meituan's technical talk on advertising algorithms, describing the three development stages of Meituan search advertising recall: multi‑strategy keyword mining, hierarchical recall, and generative recall.

Stage 1 – Multi‑Strategy Keyword Mining : The system extracts short queries (average 2‑3 characters) and focuses on high‑frequency traffic. Offline keyword mining is performed on SPU data, followed by online exact matching. Early versions relied on rule‑based extraction, then progressed to sequence labeling models, pointer‑combination models, and finally to generative models that can break the literal limits of keywords.

Stage 2 – Hierarchical Recall System : Introduced in 2022, this stage partitions traffic and supply into quadrants (strong intent with supply, generic intent with supply, weak supply, and no supply). It combines offline‑to‑online keyword pipelines, sparse‑vector retrieval with slot‑based inverted indexes, multi‑modal vector models, and graph‑based recommendation for weak‑supply scenarios. Personalized and generic intents are handled by successive versions of dual‑tower and multi‑objective vector models.

Stage 3 – Generative Recall : Starting in 2023, Meituan explores large‑model‑driven recall. Techniques include offline keyword generation with chain‑of‑thought (CoT) reasoning and RLHF alignment, multimodal vector recall enhanced by diffusion models, and domain‑specific LLMs fine‑tuned on advertising data. Knowledge distillation transfers CoT and hidden‑layer knowledge from large models to smaller online models.

The article also discusses practical challenges such as compute constraints, model scaling, and maintaining ROI while expanding the recall space. It concludes that generative algorithms broaden the decision space compared to discriminative methods, and future work will focus on larger models and end‑to‑end online generative recall.