The purchase begins before the product
“Help me make this room less exhausting.” This illustrative request does not name a product. An assistant could propose lighting, storage, cleaning, a quieter appliance—or buying nothing. It participates in turning a wish into a decision before a merchant receives an order.
The economic question begins there. Who hears the unfinished thought? Who defines the alternatives? Who can make an offer actionable? These functions need not belong to the same business.
The entrance can move
Shopify’s March 2026 account of ChatGPT shopping describes product discovery followed by merchant checkout, with the merchant retaining its transaction role. It is a disclosed commercial route, not evidence that every user has the same experience. [1]
That arrangement makes a strategic possibility visible: the assistant can become the place where a choice begins while another system accepts and fulfills the transaction. Keeping an order record does not guarantee receiving the next question. Influencing discovery does not guarantee profitable distribution.
Separate the rights
Unpack “owning the customer” into need interpretation, eligible offers, recommendation, offer formation, authorization, merchant acceptance, recovery and the next request. Then ask who can change each part and who pays when it fails.
A broad catalog may hide narrow eligibility. A compelling recommendation may lead to an unavailable offer. A completed payment may precede a difficult return. Each observation establishes something different. One market-share figure cannot explain the whole relationship.
Permission is not a good decision
Payment protocols help establish what an agent was authorized to do. They do not independently establish that it considered a sufficiently broad set of offers or chose well. The Agent Payments Protocol is useful to inspect precisely because its authorization problem is distinct from this welfare problem. [2]
A transaction can be valid and still be expensive or unsuitable. A useful comparison includes all-in price, fit, time, delivery, recovery and unwanted purchases. It should also allow the best answer to be waiting.
Advertising changes its address
If an assistant helps construct a shortlist, commercial influence can seek a place within that process. That does not mean all paid offers harm the customer. A genuine discount or useful information can improve a decision.
The question is whether the contribution survives the customer’s complete account of costs and alternatives. A commission-free intermediary can make a bad choice. A paid intermediary can produce a better net outcome. Incentive disclosure helps inspection but is not a substitute for measuring that outcome.
The cross-border question
A US–China lens should explain how businesses differ in access, integration and execution. It should not turn nationality into a shortcut for neutrality or control. Within each market, architectures can mix conversational discovery, platform infrastructure and merchant service.
For a supplier selling across borders, production capability creates potential value. Capturing it also requires an offer the buyer’s agent can discover, verify, purchase and obtain service for. The research question is where those practical constraints create bargaining power.
Independence through connection
An independent assistant may need commercial connections to act usefully, while preserving the freedom to reject the connected system’s preferred offer. Independence is therefore a property to investigate in access rights, incentives and behavior. Owning a model does not establish it.
Our working thesis weakens if agents gain usage without influencing repeat demand, or if integration consistently improves verified customer outcomes enough to dominate broader comparison. The next evidence should follow complete customer journeys. This essay does not present a live shopping benchmark or name a national winner.
Sources & scope
Sources checked: 13 November 2026.
Described discovery and merchant-checkout route; not a customer-outcome audit.
Authorization design; not evidence of recommendation quality or long-term reliability.