SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents

Large language model (LLM)-based web agents can navigate live storefronts, but they often collapse to a single “average buyer” policy, failing to capture the heterogeneity and distributional structure of real buyer populations. Existing personalization methods rely on hand-crafted prompt personas that are brittle, hard to scale, context-inefficient, and unable to represent population-level behavior. We introduce SimPersona, a framework that learns discrete buyer types from historical traffic and exposes them to LLM agents as compact persona tokens. Given raw clickstreams, a behavior-aware vector quantized variational autoencoder (VQ-VAE) induces a discrete buyer-type space capturing both real buyer behavior and merchant-specific population distributions. SimPersona maps each learned type to a persona token in the agent vocabulary and fine-tunes the agent on real browsing traces. At inference, buyers are assigned to learned types with a single encoder pass, requiring no retraining or store-specific prompt engineering. For population simulation, SimPersona samples buyer types from each merchant’s learned distribution, preserving merchant-specific variation. Across 8.37M buyers and 42 held-out live storefronts, SimPersona achieves 78% conversion-rate alignment, exhibits interpretable behavioral variation, and outperforms a baseline with 8 times more parameters on shopping tasks. We also release an open-source pipeline for converting e-commerce logs into buyer representations and agent-training traces.