Do shopping agents respond more strongly to price and rating cues?

ABXLAB is a "man-in-the-middle framework that transforms arbitrary websites into controllable behavioral testbeds," built to intercept and modify real web content before an agent sees it. The paper applies it to OneStopMarket, a web shopping environment, running "large-scale experiments across 17 state-of-the-art models" over interventions including authority, social proof, scarcity, negative framing, and incentives, plus manipulated prices, ratings, and product order. The primary study combines 10 interventions, 50 product pairs, and 3 conditions (no intervention, first product nudged, second product nudged) into 1,500 base configurations, with the agent choosing between two product pages. Against a human baseline run on the same binary pairs and nudges, the paper reports that people "had relatively modest shifts in choice probabilities (consistent with priors from the literature)," while agents "frequently exhibited much larger responses to the same cues."

The paper's own framing is that agents and humans share the same directional sensitivities — both move toward higher-rated, lower-priced, or socially-proofed options — but the two diverge sharply in magnitude: agents "occupy an unusual regime" where the direction of the bias matches humans while its size does not, and under matched or single-attribute conditions (rating- and price-matched pairs, profiles mentioning only one attribute) responses "collapse into rule-like patterns" rather than the probabilistic shifts seen in people. The paper contextualizes this against known human effect sizes: a Wayfair field experiment where a 0.5-star rating increase raised sales about 5% (Magnusson, 2022), primacy effects around 30% in two-item choice sets (Mantonakis et al., 2009), and nudge meta-analyses reporting single-digit average effects of 6–9% (DellaVigna & Linos, 2022) — effects the paper calls "modest on average and highly context-dependent" compared to what it measured in agents.

This sits alongside What makes people distrust AI agents they delegate to?, which finds delegation trust breaks on a handful of visible, irreversible actions. ABXLAB's finding is quieter and more pervasive: ordinary, reversible consumer choices — the kind a delegation study would never flag as a trust event — are already being bent by superficial cues at a scale no human shopper would show, with no single visible failure to trigger withdrawal. It also parallels Do language model groups mimic human group reasoning patterns?: another case where LLMs reproduce a human-observed pattern at the aggregate, qualitative level while diverging from it in magnitude and underlying mechanism.

The study is confined to one benchmark domain (OneStopMarket shopping), text-only nudges, and binary forced choice; the excerpt does not name which of the 17 models drove the largest effects or report per-model variation, and it does not test multimodal cues or larger choice sets. What it does establish is that, at least in this setting, agents are not just imperfect imitators of human choice — they are more exploitable by exactly the manipulative design patterns (review inflation, scarcity framing, discount theater) already aimed at human shoppers, at a magnitude the paper argues current consumer-protection norms were not built to anticipate.