How price sensitive is AI? Benchmarking LLM-generated willingness to pay in pricing analytics

Abstract Large language models (LLMs) are increasingly used to generate synthetic respondents for marketing research, yet their validity for pricing analytics remains unclear. This study benchmarks LLM-generated willingness-to-pay (WTP) against incentive-compatible real WTP elicited through the Becker-DeGroot-Marschak mechanism across four consumer product studies. We examine whether synthetic WTP reproduces real WTP in terms of means, distributions, price-response functions, profit-maximizing prices, and profit forecasts. The results show that LLM-generated WTP can appear economically plausible while producing materially different pricing implications. For two more hedonic products, synthetic respondents overstate WTP and imply overly optimistic demand and profit outcomes. For two more utilitarian products, results are more heterogeneous, ranging from near mean-level correspondence to substantial WTP underestimation. Across all studies, synthetic and real WTP distributions differ significantly, indicating that mean-level similarity is insufficient for validation. These findings suggest that synthetic respondents may support early-stage pricing exploration but should not be treated as direct substitutes for incentive-compatible real WTP data in price optimization.