Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer
econ.GN · cs.AI
Quick verdict
Claims pricing cues mislead AI shoppers only under costly search plus vague prompts; large-scale process-tracing evidence is credible but stimuli are narrow.
Summary
The paper asks whether retail pricing cues engineered for human bounded rationality — just-below pricing and promotional discount framing — influence commercially deployed LLMs acting as delegated shoppers. The authors introduce Tool-Lab, an adaptation of Mouselab-style information-board process tracing in which product attributes sit behind costly inspect_cell tool calls, and apply it across eight LLMs from Google, OpenAI, and Anthropic (N=16,400 sessions). Full-information baselines (conditional logit over ~1,014 Pareto-filtered coffee choice sets; Fieller ratio tests) show seven of eight models weight cents like whole dollars. Under acquisition costs with a vague 'best deal' prompt, diagnostic acquisition collapses (Study 1: 3.81 of 6 attributes; optimality 0.29 vs 0.89 at zero cost), while a specific 'lowest price per ounce' prompt largely preserves search and optimality. Bayesian multilevel moderated mediation supports H1-H3 (Study 1 IMM=-0.451, 95% CI [-0.704, -0.194]; Study 2 IMM=-0.228). Robustness checks (prompt rephrasing, price consolidation, 1-250 bag purchase volumes) and validation studies (exogenous information states: 99.3% optimal given held information; 0% arithmetic errors) attribute suboptimality to acquisition truncation, not computation failure. Boundary conditions: Luna resists truncation; Flash Lite errs even with full information. Limitations include a single product category, two-to-three alternative choice sets, artificial monetary inspection costs, and no human comparison group, so 'heuristic-like' remains an analogy.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under full information, commercial LLMs are largely normative: 7 of 8 weight cents like whole dollars (Table W2) and 6 of 8 treat discount savings like list-price reductions (Table W16).
- Costly attribute acquisition combined with a vague goal prompt truncates diagnostic search and produces heuristic-like suboptimal choices; a specific objective prompt largely prevents this (Tables 1-4).
- Suboptimality arises from information-acquisition failure, not computational failure: 0% arithmetic errors and 99.3% optimal choices given exogenously assigned partial information (Web Appendices F-G).
- Susceptibility to marketing pricing cues is governed by storefront information architecture and delegation-prompt specificity rather than immutable LLM properties.
- Under vague goals, Gemini 3.1 Pro's articulated objective shifts from unit-price to absolute price as purchase volume rises from 1 to 250 bags (38% to 84%), explaining truncation (Tables W13-W14).
Strengths
Novel measurable paradigm
Tool-Lab adapts Mouselab information boards to function-calling LLMs: attributes are hidden cells revealed via priced inspect_cell calls. Table 1's attribute-level acquisition (cents/weight/dollars) makes the mechanism observable rather than inferred from terminal choices.
Acquisition vs utilization cleanly separated
Web Appendix F assigns information states exogenously (7/8 LLMs choose the normative alternative in 99.3% of partial-information sessions); Web Appendix G audits concurrent computation reports (0% arithmetic errors across all LLMs and cost conditions; hallucinations 0.0-8.5%).
Scale and statistical care
Layered robustness
Honest boundary conditions
Weaknesses
Stimulus generality
Artificial cost structure
No human baseline
Specific prompt equals the answer key
Underspecified decoding and unintegrated anomalies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether pricing cues engineered for human bounded rationality — just-below pricing ($4.99 vs $5.00) and promotional discount framing — survive delegation to LLM shopping agents, and under what conditions. Its central reframing: heuristic-like choices in AI agents are not immutable training defects but products of the information state at choice time, jointly set by storefront architecture (attribute retrieval cost) and consumer prompt specificity. Three hypotheses: H1 (vague prompts moderate the cost-to-diagnostic-acquisition effect), H2 (diagnostic acquisition raises optimality), H3 (the indirect effect is moderated by specificity). Evidence spans two studies (just-below pricing; discount framing), two full-information baselines, robustness tests, and two validation studies (N = 16,400 sessions total, eight commercial LLMs from three providers).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do AI shopping agents fall for $4.99-style just-below pricing?
According to this study, not when all product information is freely visible: seven of the eight tested LLMs weighted cents the same as whole dollars. The vulnerability appears only when agents must pay to inspect attributes and the user's instruction is vague — then they skip cents or weight and pick the lower sticker price, which can be the worse per-unit deal.
What is Tool-Lab?
An experimental paradigm that adapts Mouselab-style information boards to LLM agents: product attributes are hidden cells revealed only through paid tool calls (inspect_cell), and the agent commits to a purchase via submit_choice. It lets researchers observe exactly which information an AI agent gathered before choosing, rather than only the final choice.
How should I prompt an AI shopping assistant to avoid bad purchases?
State your decision criterion explicitly, e.g. 'find the lowest price per ounce' rather than 'find the best deal'. In the paper, the specific objective prompt largely preserved optimal choices across all tested models, while vague prompts combined with information-acquisition costs caused most models to skip price components and choose suboptimally.
Were these experiments run on real shopping websites?
No. The studies used a simulated storefront with an explicit monetary cost per attribute inspection, charged to the delegating consumer. The authors note that testing costs in live e-commerce environments is future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extracted text omits several passages (Study 2 Tool-Lab procedure/results detail, Web Appendices E-H content); assessments of those parts rest on in-text summaries and the reproduced tables.
- Figures 1, 2, and W1 are not analyzable from the text extraction.
- arXiv preprint (2609.28372), not yet peer reviewed.
- Decoding parameters and model snapshot identifiers are not visible in the extracted text, limiting verification of session independence.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.