Two‑sided service marketplaces are shifting from deterministic request‑form intake to AI‑native probabilistic matching powered by large language models (LLMs). LLMs infer intent, preferences and latent constraints from natural language, forcing platforms to rebuild a provider‑side preference taxonomy that supports matching, search and pricing while remaining a useful signal for marketplace decisions.
We introduce an autoresearch loop that generates this taxonomy one occupation at a time and has been in production since April 2026 at a major U.S. consumer services marketplace, covering 132 occupations. Instead of a single global hierarchy, the loop treats each occupation as an independent generation problem and runs iterative propose‑evaluate‑keep refinement cycles.
Each candidate tag set is scored by a recalibrated six‑rubric LLM‑as‑judge framework, after which a panel of seven distinct personas applies weighted penalties to adjust the score; the process contains no hard vetoes.
A separate parity‑mapping stage then maps legacy request‑form Q&A pairs back to the generated taxonomy. The mapping first infers the provider attribute each legacy question was meant to measure rather than translating the question text literally, providing a coverage signal and an interface for human quality assurance.
The approach enables a smooth transition from fixed forms to flexible semantic matching, markedly improving tag coverage and matching quality.
Review