Can an AI agent buy from your store? Test the path from product page to checkout

An AI assistant can recommend a product and still be unable to buy it. It may read a product page but miss the selected size. It may add an item to a cart, then lose the cart when the shopper changes delivery options. Or it may reach a payment step without clearly showing the final amount to the person who asked.

That is why “agent-ready checkout” is not a pass-or-fail property of a single page. It is a chain of hand-offs between the catalog, the storefront, the cart, delivery and payment. A break anywhere in the chain can turn an interested shopper into a dead end.

Test a real shopping task

Choose a task your customers already ask for, such as “find a waterproof jacket under £150 that is available in medium and can arrive before Friday.” Write down the constraints before testing. The same request should work for a human using the site and for an agent using whichever interface you intend to support.

Follow the journey through these five steps:

1. Find the right item. Can the agent get from a category or search result to a product that matches the request? Does the product page state what it is, who it is for and what makes it different?
2. Choose a purchasable variant. Size, colour, bundle and region can change price or stock. Make sure the selected variant is explicit, and check that the cart contains that exact variant.
3. Keep the cart state. Change quantity, remove an item, or apply a valid discount. Confirm that the cart summary reflects each change and that a refresh or redirect does not silently reset it.
4. Calculate the actual total. Delivery cost, tax, duties and eligibility can depend on address and location. The customer needs to see those charges before authorizing payment. Do not treat the product-page price as the checkout price.
5. Confirm the order. The final step should show the items, selected options, delivery details and total. The person should understand when they are being asked to approve a purchase. A tool that merely fills the form is different from one that submits an order.

Record the URL, chosen variant, cart contents, displayed total and outcome at each step. If the task fails, the record should tell engineering whether the cause was missing product data, a broken session, a validation message or a payment boundary.

Make product facts agree across the journey

Agents need the same commercial facts shoppers do. Keep product name, identifier, price, currency, availability, variant, delivery promise and return terms consistent in the product page, feed and checkout response. When one source says “in stock” and another says “ships in three weeks,” the agent cannot safely make a confident recommendation.

Do not rely on a paragraph that says “from £89” if each variant has a different price and the page does not connect a price to a choice. Use visible labels and a clear selection state. Structured product data can add machine-readable context, but it should describe the same offer a customer sees. It does not repair a checkout that loses the selected item.

Google's Universal Commerce Protocol describes agent-facing commerce actions that include discovery and buying on Google's surfaces. OpenAI's Agentic Checkout Protocol describes checkout sessions that merchants can create, update and complete. These are useful examples of the direction of travel, but they do not mean every assistant can purchase from every store today. Availability depends on the platform, merchant participation and the integration. Read the current UCP overview and OpenAI checkout specification before choosing an implementation.

Preserve the customer's control

An agent that can navigate your site is not automatically authorized to make a purchase. Show the shopper the final amount and the material terms before an irreversible order is placed. Use a sandbox or test payment environment while building the flow. Test declines, expired sessions, out-of-stock variants and validation failures as well as the happy path.

Also check what happens after checkout. The order should have one durable identifier, retries should not create duplicate orders, and confirmation should be visible to both the shopper and the merchant. If your current integration cannot safely support those guarantees, stop at a pre-filled checkout and let the customer finish it.

What a website scan can tell you

An agent-readiness scan can find public-page and access issues that make the first part of this journey harder. It cannot prove that a private cart, payment provider or fulfillment system will complete a real order. Use the scan to identify technical obstacles, then run a controlled end-to-end test against your own checkout environment.

Dedica's free website readiness scan is a useful starting point for checking the pages and routes an agent can reach. Treat the result as one part of the commerce test, not a checkout certification.

If you are deciding whether a commerce integration is needed, compare the roles of UCP, ACP and x402 before choosing a build path.

Sources

• Getting started with Universal Commerce Protocol, Google for Developers.
• Agentic Checkout specification, OpenAI Developers.
• Google product structured data, Google Search Central.