LLM-Based Customer Digital Twins for Conjoint Analysis

A new framework uses large language models (LLMs) to create 'customer digital twins' (CDTs) as virtual respondents for conjoint analysis. Researchers identified active Reddit users, aggregated their review histories into vector databases, and employed retrieval-augmented generation (RAG) with prompt engineering to build agents that retrieve past preferences. These CDTs performed pairwise comparisons on product profiles from fractional factorial design; logistic regression estimated part-worth utilities. The approach aims to reduce time, cost, and respondent fatigue in market research.