In plain words
The paper introduces price-equivalent auditing, a method for measuring how much a seller’s product description can make an AI shopping agent tolerate a higher price.
The researchers raise the price of a product with a particular cue and track when the agent chooses it only half the time, while balancing which product appears first.
For GPT-4o comparing headphones, six cues all won at least 95% of equal-price choices, but their price-equivalents ranged from 1.6% to 21.4%. So an equal-price win rate can hide large differences in how much extra price a cue offsets.
The research
If an AI shopping agent recommends products, seller-written cues may shift its choice even when prices are unchanged. But when several cues win almost every equal-price choice, how can an auditor tell which one offsets a little extra cost and which offsets much more?
The authors introduce price-equivalent auditing, which measures the price premium a seller’s cue can offset before an AI shopping agent chooses the product half the time.
The authors raise the price of a cued product in a balanced two-product choice task and find the premium where the agent chooses it half the time. This price-equivalent (PE) averages choices across products, cue placements, positions, and generation randomness.
For GPT-4o on headphones, cues with near-ceiling equal-price win rates had widely different PEs; scarcity also ranked below gain/loss framing by PE despite winning more often at equal prices. The sweep therefore measures a different feature from the win rate: how much extra price a cue offsets.
These results come from controlled choices between two products. They do not establish how people or agents behave while shopping live, or whether the findings transfer to other product domains, cues, or models.
Analysis
Deep
- Original paper authors
- Felix Buecheler, Gerardo Berbeglia, Mingyu Guo
- Submitted to arXiv
- License
- CC BY 4.0