AI Can Fill the Cart. Can Your Customer Finish?

A better entrance still needs a working checkout

WebMCP gives a website a way to expose structured tools to an agent, such as product search, rather than requiring it to interpret every button. Google's documentation, updated October 7, describes a proposed standard and a Chrome origin trial. The October 27 Community Group draft explicitly says it is neither a W3C Standard nor on the W3C Standards Track. A merchant announcement should not be read as universal browser support. [2][3]

Commerce describes a flow in which the agent finds an item, adds it to a cart and returns a checkout link. For a retailer evaluating the feature, that makes the immediate question concrete: what will the customer see when that link opens? The announcement does not answer how well every store's particular catalog and checkout will perform. [1]

Consider a hypothetical outdoor retailer. A customer asks an assistant for a waterproof jacket in medium, suitable for a weekend trip. The assistant prepares a cart. Now the customer wants to check the delivery date, change the color and see the final total on a phone. This invented example is a proposed test case, not a report of BigCommerce behavior.

The elegant demonstration is the jacket arriving in the basket. The useful demonstration includes what happens next. A customer should not need a second expedition to discover the delivery charge.

Follow the handoff, not just the traffic

External evidence makes this transition worth examining. In an October 21 analysis, CHEQ followed activity on 14 retail sites from August 19 to October 18. It reported that in-browser agents handed control back more frequently on basket and checkout pages than elsewhere. It also found different conversion patterns for AI-referred visits on mobile and desktop. Those are different groups: a person following an AI recommendation is not the same as an agent browsing for that person. [4]

CHEQ sells analytics and agent-management services. Its sites were selected partly for minimum AI-referred visits and orders, and its agent samples were smaller. The findings are observational, use its own classification and do not establish that AI caused purchases, explain every abandonment, or validate Commerce's offering. They support investigating the journey, not borrowing a conversion forecast. [4]

For our fictional retailer, I would compare where ordinary shoppers, identifiable AI-referred shoppers and identifiable agent-assisted sessions stop, separating mobile from desktop. Record unknown traffic as unknown. A growing cart count can justify further investigation, but it cannot tell the commercial director whether the customer reached an understandable purchase decision. This is my application of CHEQ's measurement guidance. [4]

Let the customer recognize the purchase

The human part has its own evidence base. Baymard Institute's April 23 checkout guide recommends explicit next-step labels, prominent guest checkout, preserved information when users go back, and a review stage that summarizes known facts rather than introducing unexpected fees. This is commercial usability guidance drawn from its broader research program, not an experiment on this new agent integration. [5]

Applied to the jacket scenario, I would want the handoff to make the chosen size, color, quantity and costs easy to inspect, with a clear route to change them. If delivery information still depends on a destination, say so. The objective is not to promise certainty before the store has it. It is to let the person understand what is settled and what still needs an answer. That is an editorial application of checkout clarity, not a claim that WebMCP supplies these details automatically. [5]

This suggests a practical ownership choice: the executive responsible for ecommerce conversion should own the whole journey, with merchandising, customer service and engineering contributing. Otherwise, a technical team could declare the connection complete while nobody has checked whether a customer can change the jacket. I propose that assignment of responsibility; the cited research did not test an organizational chart.

Take one ordinary order through an awkward afternoon

Google's August 26 design guidance recommends defining the user's goal and starting state, role-playing the interaction, accounting for ambiguity and making failures recoverable. It also says production work needs real-world telemetry beyond role-play. That is platform-developer guidance, not independent proof of sales impact. [6]

My proposed first step is a short rehearsal using one product family and the actual browser and agent combination under consideration. Start with the ordinary jacket purchase. Then change one condition at a time: the requested size is unavailable, the customer edits the cart, or a sign-in step interrupts the journey. Observe what the customer and store actually receive. These are adaptations of Google's design guidance and Baymard's checkout principles, not a validated conversion method. [5][6]

• Can the person recognize and revise the order? Check the chosen variant, quantity, costs and the distinction between reviewing and placing the order. [5]
• Can the journey recover? Check whether a correction or missing detail leads to a useful next step without losing the customer's work. [5][6]
• Can we see what happened after handoff? Examine completion and abandonment by identifiable journey type and device, preserving classification uncertainty. [4][6]

Keep the customer in the definition of success

The fictional retailer might discover that its best next investment is a clearer mobile cart, better variant information or an easier correction path. It might find that the tested combination is not ready for a wider rollout. The rehearsal does not guarantee an answer; it gives the team a concrete experience to investigate before making a larger commitment.

Commerce's announcement makes the storefront a timely place to examine agent-assisted buying. Keeping checkout on your site creates an opportunity to serve the customer directly. Make that final stretch worthy of the convenience that brought the customer there.

Questions for executives

1. Who owns the journey from an agent-prepared cart to the customer's completed order?
2. Can customers review and correct the purchase without starting again?
3. What evidence would justify expanding the tested shopping journey?