Is Your Website Ready for AI Agents? A Technical Readiness Checklist

Your Website Has a New Kind of User

Browser agents can research, compare, fill forms and complete multi-step tasks on behalf of a person. They do not experience a website exactly as a human does.

Google's web.dev guidance explains that agents can combine screenshots, raw HTML and the accessibility tree. A visually polished site may therefore become difficult if its controls are custom, labels are missing, layouts shift or important terms exist only inside images.

Agent readiness begins with the web's established foundations. It is not a reason to rebuild the site around a speculative protocol.

1. Use Real Interactive Elements

Use an anchor for navigation and a button for an action. Avoid clickable divs and spans.

Native elements expose role, keyboard behaviour and state more reliably. When a custom control is unavoidable, implement the correct semantics and interaction completely; adding role="button" alone does not reproduce button behaviour.

Audit every primary task:

• navigation links
• search
• filters
• tabs and accordions
• form submission
• cart and checkout
• account actions

2. Give Every Control a Stable Name

An icon-only control needs an accessible name. A form input needs a programmatically associated label. Two buttons called “Learn more” are ambiguous when removed from visual context.

Prefer names that survive extraction:

• “View web development service”
• “Download AI policy template”
• “Submit contact enquiry”

Do not generate random labels or change control names during loading unless the state change is meaningful and announced.

3. Keep Layout and State Predictable

Agents using screenshots can be disrupted by layout shifts, overlays and moving controls. Reserve image space, avoid surprise pop-ups and keep primary actions in consistent positions across related pages.

Every state should be explicit:

• loading
• disabled and why
• validation error
• success
• empty result
• authentication required

A spinner that disappears without a confirmation message leaves both people and automation uncertain.

4. Publish Important Facts as Text

Prices, availability, specifications, deadlines, eligibility and terms should not exist only in an image, canvas or animation.

Use clear visible text and appropriate structured data where supported. Keep the values consistent across page, markup, feed and checkout.

For comparison tasks, use real headings, lists and tables with meaningful labels. A collection of styled divs may look identical while providing a much weaker machine-readable structure.

5. Make Forms Understandable

Connect every label with its input. Use suitable input types and autocomplete attributes. Explain format requirements before submission and attach errors to the relevant field.

Do not rely on placeholder text as the only label. Keep error messages specific: “Enter an email address in the format name@example.com” is more actionable than “Invalid input”.

For multi-step forms, expose progress and preserve entered information when validation fails.

6. Protect Consequential Actions

Agent-friendly does not mean frictionless at any cost. Deleting data, sending a message, accepting terms or paying money requires clear confirmation.

Before the final action, show:

• what will happen
• who or what is affected
• price and recurring terms
• editable choices
• cancellation or recovery path

Use authentication, rate limits, fraud controls and audit records according to risk. Never weaken security merely to make automation easier.

7. Provide Crawlable Navigation

Important pages should be reachable through ordinary links with meaningful anchor text. Provide stable URLs, sensible canonicals and a sitemap.

If essential content loads through JavaScript, test the rendered result and failure states. Google can process JavaScript, but complex rendering adds more places for discovery or interaction to fail.

8. Test the Accessibility Tree

Inspect the accessibility tree in browser developer tools. Confirm that:

• headings form a logical outline
• landmarks identify navigation and main content
• controls have roles, names and states
• hidden content is not unexpectedly exposed
• error and success messages are detectable

Then complete the journey using only a keyboard and a screen reader or accessibility testing tool.

9. Test Real Tasks End to End

Do not score the home page and call the website ready. Choose the three tasks that matter commercially, such as:

1. find a suitable service and submit an enquiry
2. compare products and complete checkout
3. find support information and contact the right team

Run each task at mobile and desktop sizes, with realistic validation errors and slower loading. Record exactly where the agent or person becomes uncertain.

10. Monitor and Contain

Agent traffic can create new operational patterns. Monitor errors, unusual request rates, form abuse, abandoned steps and repeated failed actions. Add controls proportionate to the risk without blocking legitimate accessibility tools or search crawlers indiscriminately.

Emerging standards such as WebMCP and Universal Commerce Protocol may create new integration options. Treat them as additions to a solid interface, not substitutes for one.

The Readiness Scorecard

Score each critical journey from 0 to 2 for semantic controls, accessible names, stable layout, textual facts, form clarity, confirmation, error recovery and security. A zero in a high-impact category is a release blocker; a perfect total is not a guarantee that every agent will succeed.

The strongest agent-ready website is usually just a very well-built website: clear enough for a customer, structured enough for assistive technology and predictable enough for automation.

Official Sources

• web.dev: Build agent-friendly websites
• Google Search: Developer SEO guide
• Google Search: Optimising for generative AI features