Who decides: The consumer or the retailer? An LLM-assisted Bayesian framework for modeling purchase decisions with retailer influence

Retailers face growing pressure to personalize promotions, manage assortments, and improve returns on marketing investments while coping with heterogeneous and dynamic consumer behavior. This study develops a hierarchical Bayesian mixture model that uncovers household shopping segments from large-scale transaction data and enhances interpretability by incorporating product-level covariates (price, discounts, and popularity). To bridge the gap between statistical rigor and managerial usability, we complement model outputs with large language model–based summaries, enabling segments to be translated into intuitive shopper profiles. We evaluate the framework on the Dunnhumby Complete Journey dataset and validate results through predictive performance, robustness checks, semantic coherence, and causal inference using propensity score matching. Causal validation shows that these profiles differ systematically in price sensitivity, with wellness-oriented households responding less to discounts than staple-driven shoppers. Benchmarking against baseline recommenders further demonstrates the trade-off between predictive precision and personalization diversity. The study contributes to research on consumer segmentation, retail analytics, and recommender systems, and offers actionable insights for managers navigating competitive, data-rich environments.