A shopping agent’s final recommendation is only part of the evidence needed to assess its decision. For a retailer or consumer, an equally important question is what the agent knew when it made that choice. New experimental research puts the retrieval of essential product attributes at the centre of that investigation.
What the experiment establishes
Davood Wadi and Yu Ma’s independent preprint, submitted September 23, 2026, at 16:47 UTC, or 18:47 CEST, tests eight commercial LLMs from three providers. Tool-Lab hides product attributes behind inspection calls. Across the work, the authors report 16,400 sessions.
With costly inspection and vague goals, most tested agents omitted information needed for unit-price comparisons and made worse product choices. Specific unit-price goals generally preserved more diagnostic search. The tasks examine just-below pricing and promotional framing in coffee purchases.
The full paper defines inspection costs as experimental monetary charges to the delegating consumer, with $0 versus $10 or $50 conditions—not ordinary API fees. Choice optimality measures the lowest unit price, excluding search costs. Results vary across models and conditions. Small, static choice sets do not establish manipulation in live storefronts or overall consumer welfare. Peer review is not confirmed.
The commercial question: which information is available at commitment?
Our editorial interpretation is that agent-facing product information deserves its own quality review. A page can contain all the necessary facts somewhere while still making the comparison difficult to inspect. For a future audit, the useful question would be whether the relevant values are available together, refer to the same variant and retain their units when extracted.
Consider a hypothetical comparison between a 250-gram package costing €4 and a 500-gram package costing €7. The first has the lower checkout price; the second has the lower price per kilogram: €14 rather than €16. Neither is the universally correct purchase. A shopper who needs only a small quantity may reasonably choose differently from one explicitly seeking the lowest unit price.
This example is our illustration, not a result from the paper. It shows why an evaluation needs a stated objective before labelling a choice a failure. Delivery charges, stock availability, dietary requirements and waste can matter to a real purchase. A benchmark that omits those factors should identify what its score actually represents.
A practical audit proposal for product teams
We propose starting with a small catalogue sample whose price, quantity and variants can be independently checked. Record a dated reference for each item. Identify which fields are necessary for the task, then verify that the agent’s extracted values match those references before evaluating its recommendation.
Keep price and quantity attached to the same product identifier throughout the record. A correct price from one variant combined with the weight of another creates a false comparison even if both numbers are individually accurate. Note how bundles, promotional conditions and measurement units are represented, and preserve the evidence used to calculate each comparable value.
For a controlled test, compare the current presentation with a version that makes the required fields easier to retrieve. Keep the products and commercial terms constant. Capture tool calls, returned fields, failures and the final choice. Repeat observations and report any missing information separately from arithmetic mistakes. This is a proposed testing method; we have not run it against a store.
Measure the cost of getting a better answer
A deployment test should also report actual operating costs and elapsed time. An agent that retrieves every field may make a better product comparison while consuming more resources than the purchase justifies. Those costs need their own accounting, alongside the quality of the selected item, rather than disappearing inside one success score.
For NetContentSEO, a useful research agenda is to examine how product facts survive the journey from catalogue to extracted record to recommendation. That includes detecting inconsistent units, missing conditions and mismatched variants. It would allow a team to distinguish a content problem from a retrieval problem or an evaluation problem.
The paper supplies a reason to investigate this chain carefully. It does not supply a guaranteed recipe for influencing AI purchases. The next step is measurable work on real product representations, with a clear buying objective and an explicit record of both decision quality and the resources spent reaching it.