Agent recommendations need conflict disclosure and outcome-level tests, not only truthful product descriptions.
Why this question matters
A recommendation can be factually correct and still systematically favor an affiliated seller. When discovery, ranking, conversation, and checkout happen inside one automated workflow, the user may never see the alternatives or commercial relationships that shaped the result.
This is a relationship problem rather than a single-output hallucination. Monitoring should compare eligible offers, ranking consistency, commissions, ownership, and the long-run distribution of selected merchants.
Signals worth observing
- Affiliates are selected more often after controlling for price and fit.
- The agent requests narrower searches when preferred inventory is available.
- Explanations omit material provider or revenue relationships.
Practical control direction
- Disclose conflicts in machine-readable and human-readable form.
- Audit selection rates against comparable alternatives.
- Separate recommendation objectives from marketplace revenue incentives.
AgentCollusion lensThe beneficiary behind an agent action is as important as the agent identity that performed it.
