8 Questions That Reveal an Agent’s Recommendation Coverage and Incentives

1. Who benefits financially from the recommendation?

Record disclosed referral commissions, sponsored placements, merchant-funded offers or platform relationships when evidence exists. Do not infer a hidden arrangement solely from an unfamiliar ranking. A commercial relationship can coexist with a suitable recommendation. The relevant test is whether the buyer understands it and whether the ranking still follows the declared criteria.

The FTC’s endorsement guidance provides a US example of why material relationships need clear disclosure. Applying that guidance does not establish global compliance. For the DN record, save the disclosure the buyer actually saw, where it appeared and whether it explained the relationship in plain language.

Record the relationship the available evidence actually supports: a referral arrangement, paid placement, merchant-funded offer or ownership relationship. Identify who pays whom and the event that creates payment where disclosed. If that information is unavailable, label it unavailable instead of inferring a hidden fee from the ranking alone.

A subscription paid by the buyer does not, by itself, prove that no other commercial relationship exists. Equally, a disclosed commission does not establish an unsuitable recommendation. The buyer needs a comprehensible account of the incentives and a separate assessment of whether the offer meets the job.

2. What buyer objective is the agent optimizing?

Freeze the intended outcome before viewing the ranking: lowest qualifying total, dependable arrival, low ongoing commitment or another declared priority. State the mandatory conditions that disqualify an offer and the preferences that order the remaining choices. Those are different parts of the decision.

Ask whether the recommendation explanation gives the buyer enough information to reproduce the decision. A statement such as best overall may conceal a trade between price and delivery time. A useful explanation names the requirement, the evidence and the accepted trade. If the system cannot provide that detail, the appropriate next step may be a narrower recommendation or a human comparison of the unresolved offers, not a confident declaration that the entire market was considered.

Do not let a label such as best overall substitute for the missing tradeoff. If a higher-priced supplier wins because of an earlier arrival or stronger support route, the record should identify that benefit and the buyer’s reason for valuing it. If no supported reason is available, the outcome remains unexplained rather than automatically commercially biased.

Keep objective changes versioned. A later instruction to prioritize speed may justify a different recommendation, but it should not rewrite the rationale attached to the earlier cheapest-qualifying-offer result.

3. Which suppliers could qualify for this specific purchase?

Define an investigated supplier set narrow enough to support evidence: named sellers able to serve the buyer’s region, required product and material terms. Every seller online is usually an unsupported denominator. Start with the buyer’s requirements and preserve how the candidate list was assembled.

Document how the supplier set was assembled. A list taken only from the agent’s own marketplace cannot independently establish what that marketplace omitted. A buyer-approved supplier roster, a scoped industry directory or a documented manual discovery process can provide a different starting point, but each has its own coverage limits. Preserve those limits rather than calling the roster complete because it contains many names. Suppliers discovered later can be added as a new version without rewriting the original denominator.

OpenAI’s shopping help explicitly describes limits in coverage and price presentation. That is a documented product boundary, not evidence that any particular omitted offer would have been better. A recommendation audit should preserve both facts: coverage may be incomplete, and outside candidates still need the same suitability checks.

Use the same eligibility rule for partners and non-partners. A cheaper seller with an unresolved invoice, warranty or delivery condition belongs in the unresolved group until the condition is checked. Excluding it as unsuitable without evidence makes the observed coverage look more complete than the record supports.

4. Which eligible suppliers are absent from the shortlist?

Classify each investigated supplier as represented and eligible, investigated and unsuitable, eligible but absent, or unresolved. These categories must be mutually exclusive at the chosen time. A seller with an unavailable warranty territory is different from an eligible seller whose feed was missing. Unknown shipping or account eligibility stays unresolved.

Suppose an illustrative audit investigates ten named suppliers. Four appear in the recommendation and qualify, two are documented unsuitable, two qualify but are absent, and two remain unresolved. The known eligible set contains six suppliers, of which four are represented: 66.7% coverage of the known eligible set. This is not 66.7% market coverage.

The two unresolved suppliers prevent a final census conclusion. If both qualify and remain absent from the shortlist, represented coverage would be 50% of eight. If neither qualifies, it remains 66.7% of six. The calculator presents the known-set percentage and unresolved count, rather than silently treating unknown suppliers as unsuitable. It also checks that the four categories add to the investigated total.

Counts can still mislead when suppliers are unequal. Missing the only supplier that meets a critical delivery date matters more than missing several equivalent offers. Preserve purchase-changing conditions alongside the count. The commercial decision is whether the observed shortlist supports this buyer’s choice, not whether a supplier count reaches an arbitrary target.

5. How are the suitable candidates ordered?

Record the stated ranking factors and the evidence of their application. A low unit price, fast-delivery badge and large customer rating are not interchangeable measures of the buyer’s objective. The explanation should connect each material advantage to a condition the buyer actually values.

Keep observed order separate from inferred cause. A candidate can move because the available offer changes, an account becomes eligible or a source is refreshed. Save the relevant state and timestamp before claiming that a particular feature drove the result. An unexplained ranking is a useful finding without attributing a motive the evidence cannot establish.

If the order changes, describe what changed and what remained uncertain. A rank difference could reflect stale availability, missing fees or a personalization setting. Repeat matched requests where permitted and keep the source records. The defensible conclusion may be that the recommendation is not reproducible enough to explain, rather than that a specific payment caused the rank.

A defensible explanation also names the comparison boundary. State whether the ranking covers the investigated set, the accessible marketplace or another defined scope. A precise first-place position does not expand that scope.

6. When can a commercial or operational exception change the result?

Create a counterfactual using the same requirements and observation time. Add one suitable non-partner, remove a sponsored candidate from consideration, or hold commercial status aside while reviewing documented price and terms. This is a proposed evaluation design. It must respect the platform’s supported interface and cannot reveal an internal ranking system merely by observing a few outputs.

A useful counterfactual also preserves the buyer’s constraints. Removing a partner seller and adding a non-partner with an unsupported return policy would not test the effect of commercial status; it would change suitability at the same time. Compare candidates that meet the same material requirements, then describe the remaining differences. If there are no suitable non-partners in the investigated set, report that boundary. It is neither proof of independence nor evidence that the agent deliberately excluded them.

Write the exception being tested before comparing outputs. A sponsored placement, a default seller or an availability fallback can alter what the buyer sees, but those mechanisms require their own evidence. If the interface does not expose the needed state, retain the limit of the observation.

Do not turn a few matched responses into a reconstruction of an internal ranking model. The strongest supported conclusion may be that a recommendation changes under a documented condition, with the cause still unresolved. Keep this distinction visible when deciding whether the buyer needs an independent comparison.

7. What missing evidence could change the recommendation?

Name the purchase-changing question and the source that could settle it. Confirming an invoice requirement, resolving an arrival window or establishing account eligibility is more useful than asking the buyer to check everything again. Assign the question to the party able to answer it and preserve the status until evidence arrives.

In the ten-supplier illustration, two cases remain unresolved. If both later qualify and remain absent, the represented share changes from four of six known eligible suppliers to four of eight. That conditional result is limited to those two cases. It does not estimate suppliers outside the investigation.

Maintain the record only for the decision window it serves. Price, availability and seller terms can change; a historical audit is still citable if its timestamp and limitations survive. A follow-up should update the evidence and explain the changed recommendation. It should not silently replace an unresolved old record with a newly confident answer that appears to have been true all along.

If the missing condition cannot be resolved before approval, show how that uncertainty affects the choice. The buyer can approve a bounded exception, choose an evidenced alternative or defer the purchase. The agent should preserve which decision was made.

8. What does the buyer need to do next?

When a recommendation includes affiliate links, preserve the distinction between editorial suitability, tracked referral and payable commission. A useful disclosure does not guarantee attribution, and working attribution does not establish suitability. The buyer should be able to understand the commercial route before deciding whether to follow it.

Provide a usable handoff: the recommended qualifying offer, material tradeoffs, commercial disclosure, unresolved questions and the next action. The action may be to verify one condition or to open a merchant checkout; those are different levels of commitment. Keep purchase authority separate from the recommendation.

The Supplier-Coverage Map above is a count of the entered, investigated set. It cannot discover omitted sellers, confirm eligibility or prove why a ranking occurred. A completed card is therefore the beginning of an explainable buyer decision, with supporting references and visible gaps, rather than a universal marketplace score.

A maintained follow-up should record what the buyer ultimately selected and what evidence changed the choice. That connects the recommendation to the decision without pretending that a referral click or a supplier mention is confirmed commercial value.

Frequently asked questions

Primary sources

1. Shopping with ChatGPT Search. Shopping output does not include all products; displayed first merchant price need not be lowest, and merchant changes may lag. Reviewed 2026-11-11.
2. FTC: Endorsement Guides questions and answers. US endorsement guidance requires material commercial relationships to be disclosed clearly; affiliate-link wording alone may not communicate the relationship. Reviewed 2026-11-11.

Related DN research

• AI Shopping Agents vs Human Buyers
• Agentic Procurement Benchmark 2027