Checkout Just Became Something Agents Can Operate

Shopify's September 28, 2026 announcement opened checkout to browser-based AI agents through WebMCP. It named three tools: get_checkout, update_checkout and complete_checkout. TechCrunch reported the change.

WebMCP is a proposed way for a web page to tell an AI agent in the browser which actions it offers, by name. The agent can call those actions instead of guessing at buttons and fields. For any site that takes payments or bookings, the practical question is which actions an agent may take and which need a person.

I build for the web and advise companies on how people and AI find and use their sites. My work includes measuring whether AI tools recommend a company and deciding how an agent should interact with its site.

What did Shopify change about checkout?

Shopify opened checkout to browser-based AI agents with three named tools: get_checkout, update_checkout and complete_checkout. The agent completes the purchase with the buyer's consent. A Shopify product manager told TechCrunch, "Shopping with an agent shouldn't feel like watching paint dry."

Those tool names make the available actions legible to an agent. A page that exposes named actions gives the agent a defined route through checkout, rather than asking it to infer which button or field comes next. That difference matters to a person responsible for the purchase flow: the action names can be reviewed as part of the design.

The product manager's comment points to the experience Shopify is aiming for. The agent can move through the checkout actions while the buyer remains part of the decision to complete the purchase. For another business, the same design question applies to its own flow: where can an agent act, and where should the person decide?

What should a non-Shopify site decide?

A site outside Shopify can start by writing down one real customer flow and assigning each step to either the agent or a person. In a booking flow, the team might let an agent gather the booking details and update the selection, then require the person to approve the booking before it is completed. The exact boundary depends on the actions the site offers and which ones the business is willing to let an agent take.

That example makes the decision specific to the flow. A booking and a purchase may have different points where the person needs to step in. Mapping the steps gives the site team something concrete to review before an agent is allowed to call them.

I published a rubric for scoring a site's readability by agents, so any firm can score itself and get the same answer.

How can I tell whether AI tools are choosing my business?

Deciding which actions an agent may take is one part of preparing a site for AI use. I also measure whether AI tools recommend a business when people ask real buying questions. That gives me a separate way to check how the business appears to AI tools, rather than treating an agent's ability to use checkout as proof that customers will find it.

My measurement uses a locked library of a category's real buying questions. I ask them of ChatGPT, Claude, Perplexity, Gemini and Grok on a fixed schedule, check official domains so a same-named company does not count as a win, and ask each model why it chose what it chose. I keep the records in one format so week 11 compares cleanly with week 1.

Over 11 weeks, Soapbox's AI recommendation win rate rose from 6.2% to 36.2% in one engagement. Trinzik did the work using the platform I built. This is directional evidence, not a controlled study. In another program, I published into the gaps, and the win rate remained flat. I retest rather than call that a victory.

Common questions