Meituan Search 3.0 rebuilds ranking with LLM semantic representations

On October 3, 2026, the Meituan technical team published a blog post on how LLM semantic representations are being applied to the ranking models behind Meituan Search 3.0. According to the post, the Meituan search team is rebuilding its local-life search foundation on a new architecture that merges the group-buying and vertical business lines, together with generative large models. Meituan Search 3.0 is presented as a broader effort to reconstruct that search foundation, and the post is one installment in a series the team says will keep documenting the technical exploration behind it.

The focus of this installment is the service-retail ranking scenario, where the use of LLM semantic representations is laid out in three phases. Phase one was single-point feature validation, checking individual signals on their own. Phase two moved from those isolated checks to building a systematic representation system. Phase three covers cross-scenario transfer and reuse. Across all three, the team describes the goal as exploring the path by which semantic matching signals can be applied to search ranking.

The post does not give start or end dates for the individual phases, and it does not report evaluation metrics, model names, parameter counts, or production results. As published, the public information stays at the level of the three-phase framework and the direction the team says it is pursuing.