An agent-friendly website is prepared for visitors that are not people: AI agents that carry out tasks on the web on a person’s behalf, such as comparing offers or booking an appointment. Such an agent doesn’t just read text; it clicks buttons, fills in forms, and moves between pages. A guide titled “Build agent-friendly websites” on web.dev, Google’s developer platform, describes how websites should be built for this. Google also calls such a website “agent-ready.”
How AI agents perceive a website
Agents don’t simply look at a website the way people do. According to the web.dev guide, they work with a machine-readable representation of the page, which they get in three main ways: from screenshots analyzed by a vision model, from the page’s structure in the browser—the Document Object Model (DOM)—and from the accessibility tree, a simplified version of the page that lists every relevant element, such as each button, with its role, name, and state. The guide notes that analyzing screenshots can be slow and costly, so it works better as a fallback when a page’s structure is confusing. From the DOM, for example, an agent infers that a “Buy now” button belongs to the product whose section it sits in. Google’s documentation for website owners names the same three channels. Modern agents combine them, so an agent-friendly website sends clear signals through every one of them.
The building blocks
What makes a website usable for agents can be sorted by whether an agent can reach a page, understand it, and operate it, and whether the site explicitly offers actions to it:
- Reach: The agent is allowed to fetch the page. Bot management can turn agents away as well, for example in a firewall, in a hosting provider’s settings, or in a content delivery network (CDN), a service that delivers websites through many servers. Cloudflare, a CDN provider, treats agents that visit a page on a person’s behalf as a category of their own with a separate setting.
- Understand: Semantic HTML marks up headings, links, buttons, and form fields as what they are. The web.dev guide recommends real button and link (a) elements over generic containers such as div or span that are only styled to look like controls, because agents recognize the former as interactive elements. Every control also needs a clear name, such as a label that is linked to its form field in the code. An llms.txt file adds a compact overview of the most important content.
- Operate: Everything needed to complete a task is visible on the page and not covered by transparent overlays, because an agent’s visual analysis may skip covered elements. The layout stays stable: elements don’t move while the page loads; otherwise, an agent may miscalculate where a button is. Recurring buttons such as “Add to cart” sit in the same place on comparable pages.
- Offer actions explicitly: Optionally, a website can use WebMCP, a proposed web standard, to tell an agent directly which actions it offers, such as booking an appointment, and what each input field means. According to Google, WebMCP is in active development and may change.
Google’s auditing tool Lighthouse checks several of these building blocks automatically in its experimental Agentic Browsing category. Progress is best recorded there audit by audit, since the number of audits can change with new Lighthouse versions.
Overlap with accessibility
Many of the building blocks come from web accessibility, because agents rely on the same roles and names as assistive technologies such as screen readers. Building accessibly thus lays much of the groundwork for agents at the same time. Conversely, according to web.dev, everything its guide suggests for agent-friendly websites also makes sites better for people.
What it means for GEO
The criteria for agent-friendly websites are young. Google describes the Lighthouse category and its WebMCP checks as experimental and based on proposed standards, noting that the standards for the agentic web are still emerging. In its documentation for website owners, Google points businesses for which agents are relevant and that have time to spare to the web.dev guide.
For GEO, agents can shift what visibility means: when agents carry out tasks, it matters not only whether an offering is mentioned in an AI answer but also whether an agent acting on a person’s behalf can use its website. Whether an agent-friendly website leads to more mentions or AI citations is not established; Lighthouse measures how well a website is built for agents, not its AI visibility. What speaks for the building blocks is that most of them are part of well-built, accessible web development and therefore make sense even while the standards for agents are still changing. Agent-friendliness complements machine readability: machine readability is about whether software can capture a website’s content; agent-friendliness adds whether software can also operate the site.