AI shopping audit prices seller cues across 113,472 trials

Felix Buecheler, Gerardo Berbeglia and Mingyu Guo proposed a price-equivalent audit of seller influence on AI shopping agents in an arXiv preprint published Oct. 20, after 113,472 trials.

The Oct. 20 arXiv preprint by Buecheler, Berbeglia and Guo matters because it turns badges, endorsements and framing into a money-based measure across seven models, where ordinary choice-rate tests can flatten once a cue wins nearly every comparison.

The authors frame the problem as a measurement gap in generative engine optimization, or GEO, the effort to understand how seller content shapes AI-generated recommendations. The issue, they wrote, is “not whether seller content shifts an agent's choice, but how much.”

Their method raises the price of the product carrying a seller cue until the model chooses it only 50% of the time in a counterbalanced, position-controlled forced-choice test. The resulting premium is the cue’s price-equivalent, a break-even threshold meant to make different seller signals comparable in monetary terms.

When translated into price, cues that looked similarly persuasive at equal prices separated by wide margins, the authors said. Their conclusion was that choice-rate audits “substantially under-resolve seller influence.”

The audit covered seven models across families and two product categories, according to the abstract. The paper also examined Jev, which the authors describe as a decision-native model because it outputs a choice probability directly rather than generating text.

Jev’s results differed from the text models tested, the authors said. They wrote that it was “resisting most cues yet remaining susceptible to the strongest social-proof claims.”

The paper’s corporate relevance is clearest for marketplaces, AI shopping tools and search systems that may rank or recommend products based partly on seller-provided wording. If agents can be moved by presentation, developers and auditors need a way to compare that influence beyond whether one product wins a binary choice.

The preprint is not yet peer reviewed.

The authors said they released the stimuli, logs and code as a benchmark that reports each model’s cue-price profile rather than a single score. The preprint was submitted to arXiv on Oct. 19 as arXiv:2610.25843v1.

Source: arXiv cs.AI preprint, Oct. 20, 2026.