AI Autoresearch Transforms Marketplace Catalogs for Probabilistic Matching

Kartik Ravisankar, Hojat Abdolanezhad, Daniel Capo, Sang Su Lee, Shishir Dash, Vijay Anand Raghavan· September 2, 2026 View original

Key takeaways

  • Marketplaces are shifting to AI-native probabilistic matching using LLMs.
  • An autoresearch loop generates dynamic, occupation-specific provider taxonomies.
  • LLM-as-judge frameworks and multi-persona critic panels refine these taxonomies.
  • This approach improves matching, search, and pricing by inferring user intent.

Who benefits

E-commerceGig EconomyProfessional ServicesReal EstateTravel

Summary

A major U.S. consumer services marketplace has deployed an autoresearch loop that uses LLMs to generate and refine provider preference taxonomies for AI-native probabilistic matching. This system moves beyond rigid request forms, inferring user intent and preferences from natural language to create dynamic, occupation-specific attribute sets.

Two-sided service marketplaces are evolving from traditional, deterministic request forms to more advanced, AI-native probabilistic matching systems. This shift is powered by large language models (LLMs) that can infer user intent, preferences, and latent constraints directly from natural language inputs. This new paradigm necessitates a re-evaluation of how provider-side preference taxonomies are generated, as these attributes are crucial for matching, search, and pricing. A significant U.S. consumer services marketplace has implemented an innovative "autoresearch loop" to address this challenge. This system autonomously generates and refines occupation-specific taxonomies, treating each occupation as an independent problem rather than relying on a single global hierarchy. It employs iterative propose-evaluate-keep refinement cycles. Each candidate tag set is scored using a recalibrated LLM-as-judge framework with six rubrics, further adjusted by weighted penalties from a seven-critic panel representing distinct personas. A separate parity-mapping stage then links legacy request-form questions to the newly generated taxonomy, providing coverage signals and a human quality assurance interface. This approach, deployed since April 2026 across 132 occupations, infers the intended provider attribute of legacy questions rather than literal translation.

Why it matters

This represents a significant advancement in how marketplaces can leverage AI to create more flexible, accurate, and scalable matching systems, moving beyond rigid data structures to dynamic, intent-driven interactions.

How to implement this in your domain

  1. 1Assess your marketplace's current reliance on fixed-form data intake versus natural language processing for user intent.
  2. 2Investigate LLM capabilities for inferring latent preferences and constraints from unstructured text.
  3. 3Design an iterative "autoresearch loop" for generating and refining attribute taxonomies specific to different service categories.
  4. 4Implement an LLM-as-judge framework, potentially with human oversight, to score and refine generated attribute sets.
  5. 5Develop a mapping strategy to bridge legacy data systems with new AI-generated taxonomies for a smooth transition.

Original post by Kartik Ravisankar, Hojat Abdolanezhad, Daniel Capo, Sang Su Lee, Shishir Dash, Vijay Anand Raghavan

"arXiv:2609.00274v1 Announce Type: new Abstract: Two-sided service marketplaces are moving from deterministic request-form intake to AI-native probabilistic matching, enabled by large language models (LLMs) that infer intent, preferences, and latent constraints from natural langua…"

View on X

Originally posted by Kartik Ravisankar, Hojat Abdolanezhad, Daniel Capo, Sang Su Lee, Shishir Dash, Vijay Anand Raghavan on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses