When do pricing cues actually mislead AI shopping agents?

A paper testing eight commercially deployed LLMs across three providers as "surrogate consumers" finds that classic retail pricing cues — just-below pricing and promotional discount framing — rarely distort an AI shopping agent's choice when all product attributes are visible to it. The cues only become effective "when attribute acquisition was costly and the delegating goal was vague": under that combination, most of the tested LLMs "omitted diagnostic attributes, and the resulting choices resembled human heuristic decisions." The paper states its own conclusion directly: "the efficacy of marketing cues in delegated AI shopping is governed by storefront information architecture combined with consumers' instructions, not necessarily immutable LLM characteristics."

The authors build "Tool-Lab," an adaptation of information-board process tracing that places each product attribute behind a costly tool call, so they can observe which attributes an agent retrieves before it commits to a purchase. They separate "diagnostic" attributes that feed the unit-price calculation (list price, whole dollars, cents, discount percentage, weight) from "non-diagnostic" ones (packaging, country of origin, calories), and treat any choice that departs from unit-price minimization as a heuristic, resource-conserving substitution — for example, swapping the normative unit-price objective for a simpler price-minimization objective. Goal specificity, drawn from goal-setting theory, is the second lever: a "specific goal prompt" that names the objective and the diagnostic attributes "largely preserves diagnostic search and choice optimality," while a vague prompt under the same acquisition cost does not. Validation studies locate the failure in acquisition rather than computation: with information states assigned exogenously, seven of eight LLMs chose the normatively correct option in 99.3% of partial-information sessions, and no LLM made an arithmetic error when asked to report unit values.

This reframes the know/act split in Why do LLM agents remember preferences but not act on them?: there, agents hold the right preference but fail to apply it; here, agents apply the right rule faithfully to whatever information they hold, but a costly-retrieval-plus-vague-prompt environment keeps the right information from reaching the choice step at all — a failure upstream of utilization. It also sits alongside Why do accurate predictions lead to poor decisions?: both locate the source of a suboptimal AI-mediated decision in how the task is structured and specified, not in a fixed defect of the model. And it inverts the delegation direction studied in Does using LLMs actually improve strategic decision making? — there a human uses an LLM as a strategic aid; here the LLM is itself the surrogate decision-maker — yet arrives at a similar lesson, that instructions and environment shape outcome quality as much as native model competence.

The paper is candid about what its reasoning-trace evidence does not show: traces are "verbal reports of how an LLM describes its task, not direct measures of the neural network activations that produce its behavior," so the study documents what an agent acquired and which objective it articulated without establishing the internal mechanism, and it cannot rule out rival explanations such as training-data regularities or planning limits. The finding is scoped to packaged-goods unit-price choices across the eight tested LLMs and the two prompt-cost conditions described, not to agentic purchasing generally. The implication the evidence supports is narrower than "AI agents are bad shoppers": storefront design (what is costly to retrieve) and prompt specificity (what counts as diagnostic) jointly determine whether human-pricing heuristics resurface in an AI surrogate — a lever retailers and consumers can each pull.