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

Abstract: LLM-based web agents can navigate live storefronts, yet they often collapse to a single "average buyer" policy, failing to capture the heterogeneous and distributional nature of real buyer populations. Existing personalization methods rely on hand-crafted prompt-based personas that are brittle, difficult to scale, context-inefficient, and unable to faithfully represent population-level behavior. We introduce SimPersona, a novel framework that learns discrete buyer types from historical traffic and exposes them to LLM-based web agents as compact persona tokens. Given raw clickstreams, a behavior-aware VQ-VAE induces a discrete buyer-type space that captures the statistical structure of real buyer behavior and merchant-specific buyer population distributions. To provide behavior-specific guidance to LLM-based web agents, SimPersona maps each learned buyer type to a dedicated persona token in the LLM agent vocabulary and fine-tunes the agent with these tokens on real browsing traces. At inference, each synthetic buyer is assigned to a learned buyer type with a single encoder forward pass, requiring no retraining or store-specific prompt engineering. For population-level simulation, SimPersona samples buyer types from each merchant's empirical distribution over the learned VQ-VAE codebook and instantiates agents with the corresponding persona tokens, preserving merchant-specific buyer population distributions. Evaluated on $8.37$M buyers across $42$ held-out live storefronts, SimPersona achieves $78\%$ conversion-rate alignment with real buyers, exhibits interpretable behavioral variation across buyer types, and outperforms a baseline with $8\times$ more parameters on goal-oriented shopping tasks. We further release an open-source data pipeline that converts raw e-commerce event logs into buyer representations and agent-training traces.

• The paper introduces SimPersona, a framework that employs a behavior-aware VQ-VAE to learn discrete buyer personas from raw clickstreams.
• It applies a novel two-stage fine-tuning strategy to ground persona tokens for actionable e-commerce agent behavior, enhancing conversion alignment.
• Empirical results show improved clustering purity, behavioral fidelity, and realistic population-level simulation versus traditional baselines.

Learning Discrete Buyer Personas from Raw Clickstreams: The SimPersona Framework

Motivation and Background

Grounded e-commerce agents powered by LLMs have demonstrated proficiency in navigating live storefronts and performing multistep goal-driven tasks. However, current approaches often default to an "average buyer" policy, failing to capture the distributional heterogeneity present in real online consumer populations. Prior attempts at introducing personas—whether through prompt-based templates, memory modules, or textual summaries—have proven brittle, context-inefficient, and incapable of matching key population-level distributions. Existing user-modeling methods, although capable of extracting high-dimensional behavioral representations from clickstreams, have not bridged the gap between persona discovery and effective persona-grounded agent behavior.

SimPersona: Technical Approach

SimPersona introduces a framework that systematizes the learning and deployment of discrete buyer personas into agent conditioning via the following modular components:

Data Pipeline and Feature Engineering

A scalable pipeline ingests raw e-commerce clickstreams (page views, searches, cart mutations, checkouts) and enriches them through joins with product catalogs and search/query logs. The result is a set of 403-dimensional buyer-shop session vectors, comprising:

• 16 behavioral scalars: grouped by exposure/volume, engagement, funnel rates, intent, and monetary value
• Product preference embeddings: three mean-pooled 768-dimensional vectors (viewed, carted, purchased), PCA-compressed to 128 dimensions each
• Auxiliary binary masks: presence indicators for each embedding channel

Each buyer is assigned to one of five funnel strata (purchaser, checkout abandoner, cart builder, window shopper, bouncer) via rule-based cascade to facilitate stratified analyses and diagnostics.

Persona Discovery via Behavior-aware VQ-VAE

SimPersona applies a behavior-aware vector-quantized variational autoencoder (VQ-VAE) to these representations:

• Encoder: Maps each buyer's feature vector to a continuous latent
• Quantization: Latents are assigned to one of K codebook entries, linking directly to K unique trainable persona tokens
• Optimization:
  • Reconstruction loss: Feature-aligned, with group-level masking and MSE or cosine distance per feature group
  • Contrastive InfoNCE loss: Triply-filtered positive pairing (funnel stratum, product similarity, behavior similarity) prevents clusters from mixing behaviorally incompatible buyers
  • Auxiliary supervision: Three heads (engagement depth, exploration breadth, purchase intensity; each trinarized as low/medium/high), regularize codebook towards interpretable behavioral axes

Dead codebook entries are monitored and reinitialized to prevent representation collapse.

Persona Grounding via Staged Fine-tuning

To map persona tokens to actionable semantic content for the agent, SimPersona introduces a two-stage SFT protocol on Qwen3-14B-Base:

• Stage 1 (Persona Grounding): LLM backbone frozen; only persona token embeddings are fit using intent-neutral browsing traces (intents like "you are interested in skirts" without disclosing purchase intent). Ensures persona semantics must be inferred directly from aggregate action statistics rather than lexical cues.
• Stage 2 (Action-oriented Fine-tuning): Entire model unfreezed; persona tokens presented alongside explicit session intents (purchase/browse) reflecting funnel stratum. The instantiation enforces the agent's ability to fuse coarse persona-driven priors with concrete session goals.
• Trace Construction: Each SFT record includes (i) system prompt/format, (ii) user prompt with intent, persona token, progress log, and DOM tree, and (iii) assistant reply as structured action + reasoning. Traces are generated by LLM-based replays of real enriched buyer sessions on live stores and filtered for successful completion by a separate LLM judge.

Scalable Deployment and Population Simulation

Persona assignment at inference is fully feedforward: a buyer's clickstream is encoded once, quantized, and mapped to the corresponding persona token. Large-scale rollouts sample persona tokens from each merchant’s empirical codebook distribution, mirroring the observed buyer mix without retraining.

Empirical Evaluation

Clustering Quality and Behavioral Semantics

VQ-VAE clusters demonstrate:

• Stratum purity: 84.5% (vs. 78.9% for k-means), with zero incompatible mixing of high/low-funnel buyers within clusters
• Auxiliary-head coherence: <0.5% of buyers assigned to engagement-, purchase-, or exploration-incoherent tokens (vs. 66.7% for k-means)
• Separation (Cosine, CH index): VQ-VAE clusters are both behaviorally coherent and well-separated

These results affirm that persona tokens encode macro-level shopping archetypes, not merely geometric similarity in latent space.

Conversion Alignment and Behavioral Fidelity

On 8.37M buyers from 42 unseen storefronts, persona-conditioned agents achieved:

• 78%+ stratified action rate alignment (ARA) with real add-to-cart and purchase rates; outperforming all-mismatch and random-mismatch baselines by 21–24 percentage points across all funnel strata
• Direct effect sizes (Cohen's d) exceeding 2.7 for purchase intensity, with p-values $<10^{-14}$ for behavioral separation on the purchase and engagement axes
• Tokens in "High Purchase" and "High Engagement" bins lead to systematically more frequent, deeper, and sustained agent activity in simulation, consistent with auxiliary labels. Exploration binning did not translate into significant behavioral differences, likely due to compression of the underlying human signal and experimental constraints

Instruction Following and Model Robustness

Compared to a general baseline (GPT-OSS-120B, 8x larger), SimPersona:

• Matches or exceeds instruction-following in cart (91.4% vs. 81.7%) and checkout (76.9% vs. 53.1%) tasks
• Maintains deeper funnel progression across more complex session objectives
• Reduces browser-level simulation errors (StagehandTargetClosed crashes) by 66% when using persona tokens even under intent-neutral agent goals, and consistently increases navigational depth and behavioral engagement

Two-stage SFT was found essential for robust persona-grounded reasoning; a single-stage approach exhibited catastrophic error tails (up to 71% agent failure rates on some shops) and increased output corruption.

Population-level Distributional Recovery

By design, the codebook can reconstruct store-specific buyer type mixtures. On population-level tests:

• Mean Jensen-Shannon divergence of 0.054 in recovering funnel-stage distributions over all held-out storefronts, indicating near-perfect agreement
• Explained variance ($R^2$) in aggregate add-to-cart, checkout, and intent strength features $>0.87$ using only learned token distributions

This demonstrates that SimPersona enables not just stylized individual simulation but realistic, population-aligned synthetic buyer traffic.

Practical and Theoretical Implications

SimPersona operationalizes persona discovery, grounding, and large-scale assignment, providing a scalable mechanism for simulating buyer heterogeneity in e-commerce contexts. Practical implications include:

• Support for population-level A/B testing, recommendations, and storefront evaluation with synthetic cohorts faithfully mirroring observed human variability, without per-store curation
• Modular, efficient, and transferable pipeline: open-source infrastructure allows rapid adaptation to new clickstream data and merchant domains
• The framework's learning of discrete, interpretable behavioral latent codes opens possibilities for integrating multimodal (visual, layout) cues and for direct embedding initialization of persona tokens to further optimize grounding efficiency
• Limitations remain for early-stage stores lacking accumulated buyer history and in the limited dimensionality of behavioral axes captured (notably, exploration breadth)

Theoretically, the approach bridges a gap between unsupervised behavioral modeling and controlled agent policy steering, suggesting further interactions between discrete representation learning, contrastive learning, and high-level grounding in agentic LLM behavior.

Conclusion

SimPersona provides an end-to-end method for deriving, grounding, and deploying discrete behaviorally meaningful buyer personas at scale for e-commerce agent simulation. The approach achieves high conversion alignment, behavioral fidelity, and instruction following across unseen stores, while establishing clear improvements over prompt-based and clustering baselines. Its modular data pipeline and rigorous SFT procedure establish a new state-of-the-art for simulating distributionally faithful, persona-conditioned online shopping agents, with implications for scalable synthetic evaluation and agent-based systems research in dynamic consumer domains.

Reference:

"SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents" (2605.14205)