What happened
- On September 23, Davood Wadi and Yu Ma posted a study to arXiv on how language models behave when a person delegates a purchase decision to them.
- They tested eight commercially deployed models from three different providers against well-known pricing tactics: prices ending in low digits and promotional framing.
- When access to information is free, those tactics barely sway the agent. The paper is explicit on that point.
- The result changes when a cost is imposed on obtaining information and the instruction is vague: then the models skip the attributes needed to calculate unit price and choose worse, in a pattern that resembles human heuristics.
Why it matters
- The failure isn’t in the model, it’s in the design of the information. The authors say so, and that shifts responsibility to whoever builds the product page and whoever writes the instruction, not to the model provider.
- For any brand selling online in Chile or the region, this has a concrete operational consequence: if the unit price isn’t exposed in a way that is cheap to read, the agent won’t go looking for it. Product pages written for a human to skim don’t work for an agent with a lookup budget.
- And a consequence for the buyer’s side: the promise that delegating a purchase to an agent removes bias doesn’t hold when searching has a cost. The bias changes origin; it doesn’t go away.
The number
8 commercial models from 3 providers. Specific instructions preserved the optimal decision; vague ones didn’t.
Context
Automated shopping left the realm of hypothesis this year. It has already been documented how software is starting to take the buyer’s place and how different AI engines describe the same brand in incompatible ways. This study adds the missing variable: what happens when looking something up costs something.
What’s next
- Preprint without peer review, classified under general economics and artificial intelligence.
- The authors don’t announce a replication with real catalogs or with open-access models.
- No timelines announced.
Bottom line
The finding flips a comfortable expectation. The agent was supposed to read everything the human buyer doesn’t. What the study shows is that, made to pay for reading, it does exactly what the human does: looks at the big price and moves on.
Sources
- Shopping by algorithm, arXiv 2609.28372, September 23, 2026.
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
