Simulation of Human Survey Responses for Market Research

Market and user research rely on surveys to understand customer preferences. Collecting rich human responses is a costly and time-intensive process. This has prompted growing interest in using large language model (LLM)-based agents to simulate human survey responses. A key open question, however, is how different forms of contextual data shape what such agents represent and how well they reflect individual preferences. Based on data collected in a six-week longitudinal study with 13 participants, combining interviews, diary entries, and surveys, we generated eight variants of multiple LLM-based agents that progressively integrate personal contextual information and evaluated their ability to reflect participants' responses to realistic market research questions. Our results show that incorporating personal data can improve alignment with human responses, with contextualized agents achieving similarity scores up to approximately 0.78 compared to around 0.60-0.68 without participant-specific context. The findings indicate that richer contextualization strategies produce measurable but incremental improvements in how well agents approximate individual preferences.