How to make your website work for AI agents

What are the key facts about AI agents and websites?

What do AI agents actually do on a website?

They do what a person would do with a mouse and keyboard, on the person's behalf. Here is what each company says about its own agent.

ChatGPT (OpenAI). ChatGPT Work runs "its own browser, running on a separate computer in the cloud." OpenAI says it can "accomplish a wide variety of tasks on the internet by reading, clicking, and typing into web pages." Its examples include "Find and book a DMV appointment" and "Sign in to your utility account and compare plans." It asks before visiting a new website by default and asks for confirmation "before consequential actions, such as completing a booking or payment." OpenAI

Claude (Anthropic). Claude in Chrome is "a browser extension that allows Claude to read, click, and navigate websites alongside you," including "filling out forms the way a person would." Anthropic Anthropic also says that when the side panel is open, "Claude takes screenshots of your active browser tab to understand webpage content." Anthropic For developers, Anthropic's computer use tool gives Claude "screenshot capabilities and mouse/keyboard control." Anthropic

Gemini (Google). Google describes Gemini Spark as a "24/7 personal AI agent." Under the heading "Browse it, book it, buy it," Google says it can do "live web browsing to gather information across multiple sites, compare options, and even help you complete bookings," and that it's "designed to check with you before taking major actions." Google

Comet (Perplexity). Comet is Perplexity's browser. Its Assistant runs in a side panel for "asking questions and executing tasks," including "actions taken on the website." Perplexity Perplexity's admin guide says business admins can allow it to "click, navigate, and fill forms on behalf of organization members." Perplexity

How does an AI agent see my website?

Google's web.dev guide, "Build agent-friendly websites" (April 2026), says agents view a site in three main ways. Google

• Screenshots. "The agent takes a snapshot of the rendered page and uses a vision model to identify elements."
• HTML. The agent reads the page's code to see how elements are arranged and related.
• The accessibility tree. This is the browser's summary of a page for screen readers. Google calls it a browser-native API that "distills the DOM into what's most important: roles, names, and states of interactive elements."

OpenAI is specific about its own browser. In its Publishers and Developers FAQ, it says "ChatGPT Atlas uses ARIA tags—the same labels and roles that support screen readers—to interpret page structure and interactive elements," and that "Making your website more accessible helps ChatGPT Agent in Atlas understand it better." OpenAI

Perplexity says Comet's Assistant "pulls in the visible text, headlines, and website metadata from your open page" each time someone asks it something. Perplexity

So the same page gets looked at, read as code, and read the way a screen reader reads it. A page that only makes sense to the eye, or only after scripts run, fails at least one of those. Most AI crawlers that fetch pages for answers don't run JavaScript at all; see Can AI read JavaScript websites?

How do I make my buttons and forms usable by AI agents?

Use the plain building blocks the web already has. Google's guide is direct about this. Google

• Real buttons and links. Google says to "use <button> and <a> tags over modified <div> and <span> elements." If you must use something else, "always provide the element the appropriate role and tabindex."
• Labels tied to fields. "Add the for attribute on <label> tags to link them to inputs." That tells the agent a box is "Phone number," not just an empty box.
• Look clickable. Google calls cursor: pointer "a strong signal for actionability," and says interactive elements must be visible and "larger than 8 square pixels."
• Hold still. "Agents that take screenshots will likely be confused if your website layout is constantly shifting." Late-loading banners that push the form down are a problem for people too.
• No invisible covers. Google says to "Avoid transparent overlays that hide interactive elements."
• Describe what things do. OpenAI asks developers to follow WAI-ARIA best practices "by adding descriptive roles, labels, and states to interactive elements like buttons, menus, and forms." OpenAI

What information should an AI agent find without clicking around?

This part is our advice, built on how the agents describe their own work. An agent sent to "find a plumber who can come Saturday under $200" is comparing sites. Every fact it has to dig for is a chance to give up or guess.

1. Prices. Put real prices, or honest starting-at prices and ranges, in the page text. If you don't, the agent compares you on someone else's numbers. See Should you show prices to AI?
2. Hours and availability. Write your hours as text, not only inside an image or a map widget. If you take bookings, let the available times show without signing in.
3. Service area and what you do. List towns, services and limits plainly, for example "We don't do commercial jobs."
4. One stable address per thing. Give each service, product and location its own page with a URL that doesn't change. An agent, or the person it reports back to, should be able to open that exact page again.
5. Content in the HTML. Put the important text in the HTML your server sends, not only after scripts run. See Can AI read JavaScript websites?

Do CAPTCHAs and sign-in walls block AI agents?

Often, yes, and the companies say so.

• ChatGPT. "Some sites block automated browsers or require a CAPTCHA. ChatGPT may not be able to complete a task on those sites." When a site blocks access, "ChatGPT will let you know and, when possible, try another way to complete the task." OpenAI
• Claude in Chrome. Anthropic lists "Bypassing captchas" among the things Claude is not allowed to do. Anthropic
• Sign-in. When ChatGPT Work meets a login screen, "it pauses and asks you to enter your credentials." OpenAI So sign-in isn't a dead end, but every pause is a point where a busy customer may stop.

How do I make booking and checkout work with AI agents?

Two routes exist today. The agent can use your normal booking pages, or, for online stores, you can connect checkout to ChatGPT directly.

Your own booking flow. Agents ask the person before they finish. OpenAI says ChatGPT Work "will always ask for confirmation before consequential actions, such as submitting your information to book an appointment or completing a payment." OpenAI Anthropic tells developers to have a human confirm "completing financial transactions, or agreeing to terms of service." Anthropic So the agent fills in the form, then hands the final click back. Our advice: keep that path short, show the total before the last step, and make the confirmation page say clearly what was booked.

The Agentic Commerce Protocol (ACP). ACP is "an open standard for programmatic commerce flows between buyers, AI agents, and businesses," built by Stripe and OpenAI under the Apache 2.0 license. The site says "OpenAI is the first AI platform to implement ACP with ChatGPT." ACP OpenAI's docs say supporting Instant Checkout in ChatGPT takes three pieces: OpenAI

1. A product feed. A regularly refreshed CSV or JSON file with prices, inventory and fulfillment options.
2. Checkout endpoints. ChatGPT calls them to create or update an order; you validate it, charge sales tax and accept or decline it.
3. Delegated payment. OpenAI passes a one-time payment token to your payment provider. Stripe's Shared Payment Token is the first compatible one.

OpenAI says it "is not the merchant of record"; the order and the payment stay on your systems. Stripe also offers ways for sellers to "make their products and services available through agents." Stripe For products in AI shopping more broadly, see How to get products into AI shopping.

Can I block AI agents with robots.txt?

Mostly not, and blocking them usually costs you customers. Most companies treat a visit that a person asked for differently from a crawl.

Sources: OpenAI, Google, Perplexity, Anthropic.

Google says Google-Agent "is used by agents hosted on Google infrastructure to navigate the web and perform actions upon user request," and warns "The user agent string can be spoofed." Google is also "experimenting with the Web Bot Auth protocol" so sites can confirm a request really came from Google. Google

Not every agent announces itself. A browser extension like Claude in Chrome or an AI browser like Comet works inside the person's own browser, so the visit may look like an ordinary one.

Can my website give AI agents a shortcut?

Yes, but it's early. WebMCP is a proposed web standard that lets a site "make its capabilities available to an agent as a set of predefined tools," so the agent doesn't have to work out every button. OpenAI's examples include a dashboard that lets the agent set a date range and a travel planner that lets it compare options and update an itinerary. OpenAI calls its version "Site tools," and its agents can use them in the ChatGPT desktop app's built-in browser. OpenAI Google's guide says WebMCP is "currently available for experimentation in Chrome." Google

Two limits matter. OpenAI says tools defined through plain HTML form attributes "aren't available as site tools" in its browser yet; tools must be registered with JavaScript on the main page. And website tools are treated as untrusted: "normal website-access and confirmation policies still apply." OpenAI

What should be on my AI agent checklist?

How do I test whether an AI agent can use my site?

Give an agent a real task and watch it. Ask ChatGPT, Claude in Chrome, Gemini or Comet to do something a customer would do on your site, like "find the price of a gutter cleaning on this site and pick the first Saturday slot," and stop it before it submits anything. Note every place it hesitates, guesses or gives up.

Then check the page the way a screen reader does. In Chrome, open DevTools, select a button and look at its Accessibility pane: it should have a clear name and role. If it says nothing, an agent reading the accessibility tree gets nothing either.

Our AI Access scan reads every discoverable page of your site and runs 30+ checks on each one, then tells you what to fix first.

Sources

1. Browser — OpenAI. Read Oct 6, 2026.
2. Site tools — OpenAI. Read Oct 6, 2026.
3. Publishers and Developers - FAQ — OpenAI Help Center. Read Oct 6, 2026.
4. Overview of OpenAI Crawlers — OpenAI. Read Oct 6, 2026.
5. Agentic commerce: Key concepts — OpenAI. Read Oct 6, 2026.
6. Agentic Commerce Protocol — Agentic Commerce Protocol (Stripe and OpenAI). Read Oct 6, 2026.
7. Agentic commerce — Stripe. Read Oct 6, 2026.
8. Get started with Claude in Chrome — Anthropic. Read Oct 6, 2026.
9. Use Claude in Chrome safely — Anthropic. Read Oct 6, 2026.
10. Computer use tool — Anthropic. Read Oct 6, 2026.
11. Does Anthropic crawl data from the web, and how can site owners block the crawler? — Anthropic. Read Oct 6, 2026.
12. Gemini Spark — Google. Read Oct 6, 2026.
13. Build agent-friendly websites — Google (web.dev). Read Oct 6, 2026.
14. Google user-triggered fetchers — Google Search Central. Read Oct 6, 2026.
15. Google-Agent — Google Search Central. Read Oct 6, 2026.
16. Assistant Panel — Perplexity. Read Oct 6, 2026.
17. Managing Comet Assistant permissions — Perplexity. Read Oct 6, 2026.
18. Perplexity Crawlers — Perplexity. Read Oct 6, 2026.