Can an AI agent use your site, or only read it?

Reading a product page and completing an order are different tasks. An agent may find the right item but struggle with an unlabelled option, an unclear error message or a checkout step that gives no sign of what happened.

Follow one task from start to finish

Choose a journey that matters to the client: requesting a quote, booking an appointment or buying a product. Walk through every step, including a validation error and a return to the previous page.

Look for controls whose purpose is unclear. A clickable image without a useful label, a field identified only by placeholder text, or a success message that disappears immediately can make the flow difficult to interpret.

A keyboard-only pass is a useful first check. Can you reach each control, see where focus is and understand why a submission failed? It catches accessibility problems worth fixing in their own right. It does not prove that an agent will complete the task: agents use different browser tools, and some can interact with interfaces that remain inaccessible to people.

Give forms and controls a clear meaning

Use buttons for actions and links for navigation. Give fields persistent labels and explain required formats before the user submits. Associate errors with the affected fields, and make the result of an action visible.

These are established accessibility practices. The W3C forms tutorial covers labels, instructions, validation and user notifications. Much of the same structure also gives automated tools better information about a page.

Avoid treating accessibility as a proxy score for agent readiness. Test with assistive technology where appropriate, and test the actual agent journey separately if it is part of the client’s service.

Keep product facts consistent

A product’s price, currency and availability should agree across the visible page, structured data and checkout. If a promotion has ended, update each place that carries its price.

An assistant comparing suppliers may use a page extract or product data rather than complete the checkout. Conflicting values make its answer harder to trust, even when the purchase flow itself works. Check a few variants and out-of-stock items, not just the default product.

Decide what an agent may do

Allowing a crawler to read public content does not authorize it to place orders, change an account or use the content for model training. Those decisions belong in separate controls.

Crawler preferences can express access and usage choices. Transactions still need the same authentication, validation and confirmation rules as other customer actions. Do not remove those protections to make an automated test pass.

For most teams, the first useful improvement is an ordinary one: make the existing journey clear and dependable. Build a separate agent API only when a specific workflow needs it.