AI shopping agents were less likely to select the lowest unit price in a simulated test when they received a vague instruction and faced an artificial cost for inspecting product details, researchers report.
Davood Wadi and Yu Ma of McGill University tested eight large language models using a controlled shopping environment called Tool-Lab. Their working paper, revised in October 2026, has not been peer-reviewed.
The findings concern simulated choices under specified instructions and costs. They do not establish how often real customers receive poor recommendations or how commercial shopping assistants will perform in ordinary use.
Why the size of a package matters
A unit price expresses the price for a standard amount, such as an ounce or kilogram. It can make differently sized packages easier to compare. A lower sticker price is not necessarily the lower price per unit.
In one experiment, the models chose between coffee products after requesting separate pieces of information. The researchers varied whether each inspection carried an imposed cost and whether the goal clearly specified the lowest price per weight or used vaguer wording.
With a vague goal and no inspection cost, the lowest-unit-price option was chosen in 89% of sessions on average across the models. With a simulated $10 cost per inspection, that figure fell to 29%. The $10 was an experimental condition, not a claim that commercial assistants charge shoppers that amount.
With the specific goal, the corresponding averages were 94% with free inspection and 83% with the imposed cost. The models often omitted package weight and cents under the vague, costly condition. Results differed by model.
A test of information access
Tool-Lab placed information behind calls that the models had to make before choosing. That design let the researchers inspect which attributes were requested, rather than observing only the final answer.
Because separating dollars and cents into different fields is artificial, the authors repeated part of the test with price presented as one field. In that version, the tested models could still omit package weight under the vague, costly condition.
The authors also investigated promotional framing. The research does not show that a discount label will reliably persuade every shopping agent or increase a retailer’s sales.
What merchants can check
For a retailer, the work suggests checking whether product information is complete, consistent and accessible to the systems customers use. Price, quantity, package size, shipping charges and availability each affect what a buyer receives.
A comparison that omits delivery or subscription terms can be misleading even if the arithmetic on the visible price is correct. Companies should also verify that product feeds match their actual checkout terms.
Consumers can state what matters, such as a maximum budget and a preferred unit price, and check the underlying offer before authorizing a purchase. Unit price is useful, but quality, delivery and dietary or technical requirements may matter too.
MBN’s coverage of reducing dependence on search traffic examines why businesses should make information useful across more than one discovery channel.
Simulation is not a market forecast
The researchers used defined product choices, prompts and simulated information costs. Real shopping systems use different tools and may have access to information in other formats. Their performance can change as models and interfaces change.
The study offers a practical reason to test a buying assistant’s complete workflow: a recommendation can be wrong because relevant information was never retrieved, even when the calculation would have been straightforward once the details were available.
