How AI Agents Actually Use Websites (and How to Make Yours Agent-Ready)

AI agents use websites by reading a machine view of the page (a screenshot, the HTML or the browser's accessibility tree), deciding what to do next and then clicking, typing and submitting like a person would. They fail on the same things that trip up screen readers: unlabelled buttons built from `div`s, forms without labels, layouts that shift, invisible overlays and flows that hide prices or rules until late. To make a site agent-ready, fix semantic HTML and accessibility first, make key facts and policies visible, keep sensitive actions behind explicit confirmation, and consider exposing structured tools (WebMCP, an API or commerce protocols) for your most important tasks.

For two decades websites have had two audiences: people and search crawlers. A third has arrived. Browser agents (ChatGPT agent, Gemini's agentic browsing in Chrome, agent modes in AI browsers, and custom agents built on computer-use models) visit sites to finish a task: compare three insurance quotes, book a table, reorder office supplies, fill in a supplier form.

Google's web.dev team put it directly in its April 2026 guide *Build agent-friendly websites*: websites have a new type of visitor, and interfaces built around complex hover states, shifting layouts and fluid motion can be functionally broken for agents. Google's guide to generative AI features in Search makes the same point from the search side, noting that browser agents may access websites to complete tasks such as reservations.

The web.dev guide describes three views agents use, often in combination:

The accessibility tree is the closest thing a browser has to a map of what can be done on a page. It strips away visual noise and lists each control with its role (button, link, textbox, checkbox), its accessible name ("Add to cart", "Email address") and its state (checked, expanded, disabled). An agent that can read that map does not have to guess what a styled box does.

That is why so much agent-readiness advice overlaps with accessibility work. A site that already meets WCAG basics for screen-reader users is most of the way to being usable by agents. Our website accessibility guide and accessible UI design guide cover the details.

The same problems appear again and again when agents try to complete tasks on business sites:

Businesses also use the same screen-reading technology to automate their own systems that lack APIs; see computer-use agents for business.

Most of these fixes are small, improve the experience for people too, and can be shipped incrementally. Google's web.dev guidance includes several of them: semantic elements, labels linked to inputs, stable layouts, no ghost elements, `cursor: pointer` on clickable items and interactive targets large enough to detect.

Clicking through an interface is a fallback. It is slow, it breaks when the design changes, and it makes agents guess. The direction of travel is to let sites declare the actions they support so agents can call them directly.

WebMCP is a proposed web standard for exactly that. A page registers tools (a name, a description, an input schema and a function) through a browser API, and an agent in the browser calls them instead of operating the UI. Chrome shipped a developer trial in May 2026 and runs a public origin trial from Chrome 149 through Chrome 156; Chrome's documentation suggests uses such as filling complex structured forms correctly. The API surface is still changing (Chromium moved the entry point from `navigator.modelContext` to `document.modelContext`), so treat it as an experiment rather than a requirement. web.dev's guide notes that sensitive actions should still require user confirmation.

Outside the browser, the same idea already works through APIs and protocols: the Model Context Protocol for connecting AI applications to your services, and commerce protocols such as UCP and ACP for product discovery and checkout, covered in our agentic commerce guide.

Whether your business should offer such an interface, and what to expose first, is covered in should your business build APIs for AI agents.

An agent acting for a customer is still acting for a customer. Payment, account changes, bookings with cancellation fees and anything involving personal data should have a clear review step that an agent can surface to its user before proceeding. Current agents generally pause for confirmation before purchases and other sensitive steps; your interface should make that pause natural, with an unambiguous summary of what will happen.

On your side, treat agent traffic like any other client: authenticate accounts properly, rate-limit, and decide which agents you trust. Signed agent requests now make that possible without blocking everyone; see how to verify AI agent traffic.

Put together, the shift is less dramatic than headlines suggest but real. A website now has to be three things at once: a persuasive experience for people, a crawlable source of facts for search and AI answers, and an operable interface for software acting on someone's behalf. The businesses that adapt early will not be the ones with the most novel AI features; they will be the ones whose facts are clear, whose flows are accessible and whose key tasks can be completed by whatever the customer uses to reach them.

Being visible to an AI answer engine and being usable by an autonomous agent are different problems. Discoverability is about reading: crawler access, indexable pages and specific facts worth citing (see AI search visibility). Interaction is about doing: semantic controls, stable flows, visible prices and policies, authentication, confirmations and, increasingly, action interfaces such as APIs or WebMCP tools. A site can be cited constantly in AI answers and still be impossible for an agent to book or buy on, and the reverse.

AI agents read websites through screenshots, HTML and the accessibility tree, and they struggle wherever the page is ambiguous. The fix is mostly good engineering you should be doing anyway: semantic, accessible, stable interfaces with honest information up front and clear confirmation steps. Structured tools such as WebMCP are worth watching and piloting, but the accessible interface is the part every agent can use today.

For the strategic picture of how websites now serve people, search engines, AI systems and agents at once, read will AI agents replace websites?.