Why this question now
Google’s Universal Commerce Protocol, published in January, describes how shopping agents can connect to businesses and payment providers. Its stated design lets businesses keep their own business logic and remain the Merchant of Record, while cryptographic proof backs user authorization. Those are decisions about who controls the sale, even when the customer starts shopping somewhere else. SRC-001
OpenAI’s March product update expands its Agentic Commerce Protocol to support product discovery inside ChatGPT. It describes a Walmart environment inside ChatGPT with account linking, loyalty, and Walmart payments, while leaving merchants the option to use their own checkout experiences. Personalization, local availability, and ETAs are described as capabilities the protocol will support over time. SRC-002
That combination creates a more specific question than “will agents shop?” A consumer may discover a product in one system, authorize a purchase in another, receive it through a merchant, and ask for help through whichever interface remembers the order. The customer relationship becomes a chain of control points.
What the systems promise
The protocol documents are careful about the division of labor. OpenAI’s product discovery announcement describes Agentic Commerce Protocol as the connective layer between merchants and users throughout discovery. SRC-002
Its key-concepts guide says ChatGPT can render the checkout experience, while the merchant validates the order, determines fulfillment options, calculates sales tax, analyzes payment and risk signals, processes payment, and accepts or declines the order on its own systems. The order is not made real by the interface alone. SRC-004
Google’s UCP says something similar in a different vocabulary: the merchant exposes capabilities, retains business logic, and remains Merchant of Record. Visa’s merchant-facing material pushes in the same direction: an in-brand merchant assistant can help a business keep control of the experience and relationship, while a catalog and checkout made agent-friendly can let external agents shop on the website. SRC-001 SRC-005
This is the promise layer. It tells us what the builders want the architecture to preserve. It does not tell us what customers will actually remember, where they will return, or which side will have enough context to solve a problem after the purchase.
What is already being deployed
The discovery layer already has merchant integrations. OpenAI’s March announcement names Target, Sephora, Nordstrom, and other retailers as integrated with ACP for discovery. It also says Shopify merchants’ product data is integrated through Shopify Catalog without additional work by individual merchants. Those integrations put products in front of shoppers; they leave the repeat-customer question unanswered. SRC-002
On the payment side, Visa, BILL, and Billtrust announced a live B2B pilot in October 2026 in which an authorized business agent completes approved invoice payments through supplier portals. The buyer’s instructions, tokenized credentials, agent identity, and payment visibility remain part of the flow; supplier control and oversight remain explicit. SRC-006
The system needs to know who the agent represents, what it may do, which instruction authorizes the action, and what evidence remains afterward. Visa’s reported pilot tests that controlled payment handoff, rather than the question of where a consumer returns to shop. SRC-006
OpenAI’s checkout update gives merchants another reason to distinguish the sale from the customer relationship. The company says its initial Instant Checkout experience lacked the flexibility it wanted to offer, so it is allowing merchants to use their own checkout experiences while focusing on product discovery. A merchant can keep control of checkout while discovery happens inside ChatGPT. Whether the buyer returns to ChatGPT or goes directly to the merchant next time remains open. SRC-002
The three futures worth separating
Merchant keeps the front door. Agents become excellent discovery and comparison tools, but the buyer usually returns to the merchant’s account, loyalty program, service desk, and post-purchase records. In this future the agent is a high-intent referral channel. The differentiator is not the first click; it is whether the merchant can recognize and serve the buyer again.
Agent platform becomes the gatekeeper. A small number of interfaces own the repeated conversation, the memory of preferences, the ranking logic, and the next recommendation. Merchants still fulfill orders, but visibility and re-engagement are mediated by the agent. The differentiator is whether merchants can measure a continuing relationship beyond a referral and whether customers stay inside the agent surface when something goes wrong.
Control is shared by protocol. Identity, consent, receipts, loyalty, and service history become portable enough that a consumer can change agent surfaces without losing the relationship, while a merchant can change channels without rebuilding its transaction system. The differentiator is an actual switching path, not a standards announcement. These futures can coexist by category; they are a comparison set, not an exhaustive probability distribution.
What would change my read
The next useful evidence is not another announcement that agents can compare products. It is a record of what happens after the order. Who holds the account? Who can answer a return question? Can the buyer revoke the agent without losing loyalty status? Can the merchant distinguish a repeat customer from an anonymous platform conversion? Can the buyer carry the receipt and consent record to another agent?
A strong merchant-front-door signal would be repeated account linking, branded service, and loyalty participation after agent discovery. A strong platform-gatekeeper signal would be repeat shopping and post-purchase support staying inside the agent, with merchants competing for ranked exposure. A strong protocol signal would be cross-platform portability of permissions, receipts, and service context.
The record should also include the negative evidence. A merchant integration can exist without meaningful adoption. A payment pilot can prove authorization without proving loyalty. A consumer survey can measure comfort without measuring behavior. Promise, deployment, and adoption belong in separate columns.
Practical implications
For merchants, the immediate work is to test the relationship, not just expose the catalog: can a buyer return an item, recover a receipt, preserve loyalty status, revoke an agent’s permission, and reach a human service channel? The answers should be explicit for identity, consent, and the data the merchant can retrieve after an agent referral.
For agent builders, test the handoff under failure conditions: does the receipt identify the merchant and terms, can the person correct identity or loyalty mistakes, is consent revocable, and does escalation reach the right merchant team? A smoother buy button is not enough if the return or service path disappears.
For readers, the practical test is simple: after an agent helps you buy something, can you still find the merchant, the terms, the receipt, the service channel, and the permission you gave? If the answer is only “ask the same agent again,” the dependency may have shifted toward that agent for this part of the journey.
What to revisit
Next review: 2027-03-31, or earlier if a major agent-shopping launch publishes account, loyalty, service, and attribution behavior. The living record keeps the conditions and sources visible as the answer changes.
Source posture: This issue uses official product, protocol, and company pages as primary sources. Company descriptions are treated as evidence of a stated design or deployment, not independent proof of adoption or performance.