AI agents don't just answer questions; they open websites and do things. Ask an agent to "find a pet-friendly hotel in Lisbon under €180 a night and book it," and it browses booking sites, reads policies, compares options and fills in forms while you supervise. AI browsers and agent modes from the major AI companies are bringing this to mainstream users. For brands, your website is no longer only read by people and crawlers. It's also operated by software, and how easy it is to use affects whether the agent picks you.
What AI agents are doing today
Agents can already:
- Research and compare: open several product or vendor pages and summarize the differences.
- Fill forms: sign-ups, quote requests, reservations, applications.
- Complete purchases and bookings, often pausing for the user to confirm.
- Gather information from logged-in tools when the user gives them access.
Examples include OpenAI's agent features in ChatGPT and its Atlas browser, Perplexity's Comet browser, Google's agentic features in Chrome and Search, and Anthropic's Claude for Chrome. Capabilities and availability are changing quickly.
How agents read your site
Agents usually combine:
- The HTML: text, links and structure.
- The accessibility tree: the same structure screen readers use, including roles, labels and states of buttons and form fields.
- Screenshots: visual understanding of the rendered page.
That means the practices that make a site accessible to people with disabilities also make it usable by agents.
Why it matters for brands
- Agents shortlist. An agent comparing options may skip sites where key facts are hard to find.
- Agents abandon friction. Confusing forms, unlabeled buttons, pop-ups and CAPTCHAs can stop a task, and the agent moves to a competitor.
- The visit may be the conversion. An agent completing a booking is a customer, even though no human browsed your pages.
Making your site agent-ready
1. Use semantic, accessible markup
- Real
<button>and<a>elements, not clickable<div>s. - Labels on every form field, with clear error messages in text.
- Descriptive link text ("View pricing", not "Click here").
- Logical heading structure.
These are standard accessibility practices; follow WCAG guidelines.
2. Put key facts in plain text
Prices, plan limits, shipping costs, return windows, opening hours, availability and eligibility rules should be readable text on the page, not hidden in images, tooltips or PDFs. See how to write content AI assistants cite.
3. Simplify important flows
- Short sign-up, quote and checkout forms.
- Guest checkout where possible.
- Predictable steps with clear progress.
- No surprise fees at the last step.
4. Reduce interruptions
Aggressive pop-ups, interstitials and auto-playing overlays get in the way of agents and people alike. Keep them minimal and easy to dismiss.
5. Be careful with bot protection
Legitimate agents acting for users can be caught by bot defenses. Rather than blocking all automation, protect sensitive actions (payments, account changes) with standard security controls, and allow user-triggered AI fetchers such as ChatGPT-User, Claude-User and Perplexity-User. See our AI crawlers guide.
6. Keep structured data accurate
Product, Offer, LocalBusiness and Organization schema help agents and AI systems confirm facts quickly.
7. Consider machine-friendly integrations
For commerce, emerging standards such as the protocols AI companies have published for agentic checkout let assistants complete purchases more reliably. See ChatGPT shopping.
Measuring agent activity
- Server logs: look for user agents of AI fetchers and agent browsers, verified against published IP ranges where available.
- Analytics: some agent sessions show unusual patterns, such as very fast navigation or no scrolling. Don't filter them all out as bots; segment them.
- Task testing: ask an agent to complete key tasks on your site (find pricing, request a demo, buy a product) and note where it fails.
Worked example: an agent booking a demo
Ask an agent to "book a demo with a CRM vendor that supports real estate teams, for next Tuesday afternoon." Here's where sites commonly fail:
- The agent can't confirm fit. The site never says "for real estate teams" in text, so the agent picks a competitor that does.
- The demo button is an unlabeled icon or a
<div>with a click handler, so the agent can't identify it reliably. - The form has unlabeled fields or a custom dropdown that doesn't work with keyboard or accessibility tools.
- A full-screen pop-up appears halfway through, and the agent can't find the close control.
- The calendar widget only works with mouse hover and has no accessible alternative.
Each of these is also a problem for human visitors using assistive technology. Fixing them helps both.
An agent-readiness checklist
- Key facts (pricing, eligibility, policies) are in plain text
- Buttons and links are real elements with clear labels
- Every form field has a label and text error messages
- Flows work with keyboard alone
- Pop-ups are minimal and easy to close
- User-triggered AI fetchers aren't blocked
- Structured data matches the visible content
The bottom line
As agents start doing the browsing, a website has to work for software as well as people. Accessible markup, plain-text facts, simple flows, sensible bot protection and accurate structured data make your site easy for agents to understand and use. Most of it is also good for human visitors.
For how AI browsers such as Perplexity Comet and Gemini in Chrome fit in, and what ChatGPT Atlas's shutdown means, see AI browsers and SEO.
