Does higher purchase volume make AI shoppers search less carefully?

Within the same shopping-agent paper, an exogenous manipulation of purchase volume (1 to 250 bags ordered) produced a reversal of the expected direction: larger volume should raise the payoff of finding the truly cheapest option, but instead it led one of the tested Gemini models ("Pro") to abandon the unit-price objective. "The share of Pro's pre-search reasoning traces that articulated a price objective increased from 37% to 86%" as volume rose, while "diagnostic acquisition fell from 5.55 to 4.22, and choice optimality fell from .80 to .35." The pattern replicated across the three Gemini LLMs the authors tested (Table W11).

Holding volume constant, the paper reports that sessions whose reasoning trace articulated the simpler price objective rather than the unit-price objective acquired fewer diagnostic attributes and, at 1, 5, and 50 bags, were "40 to 69 percentage points less likely to select the optimal alternative." The authors read this as evidence that, under a vague delegating goal, the objective an LLM adopts — not only the cost or benefit of search — governs which attributes it bothers to acquire. A higher-stakes task did not reliably recruit more careful information-gathering; it instead appears to have triggered a quiet switch to an easier-to-compute goal.

This sharpens the acquisition-cost account in the sibling note When do pricing cues actually mislead AI shopping agents? by isolating a case where search truncation is not driven by retrieval cost at all — the volume manipulation did not change what it cost to look something up — but by objective substitution instead. It also complicates Can aggregated LLM rankings match human expert judgments?: that paper finds that scaling up the number of aggregated model judgments improves convergence toward expert rankings, whereas here scale on a different axis — the magnitude of the decision itself — degraded a single agent's performance, a reminder that "scale helps" claims about LLM judgment depend on which quantity is being scaled. It also parallels the acquisition/utilization split in Why do LLM agents remember preferences but not act on them?: the agent here, too, applies its chosen rule consistently, but the rule itself silently changes as context shifts.

The result comes from one model family in one validation study nested inside a larger pricing-cue experiment, and the authors classified reasoning traces for objective substitution only here, not across their full battery of tests, so they call the evidence for objective substitution versus value-based omission "suggestive rather than conclusive." Held to that scope, the implication is that stakes alone do not make a delegated AI shopper more careful: a vague prompt leaves room for the agent to redefine what "best" means as circumstances change, which is a distinct risk from the attribute-omission mechanism described in the sibling note.