Agent-friendly websites: understandable information and authorised actions

Start with a person’s task

Use semantic headings, ordinary links, labelled fields and factual service descriptions. A visitor or assistant should be able to identify the provider, relevant scope, conditions and next step. Where the price requires a quote, say so; do not expose an invented number merely to complete a machine-readable field. Readable HTML is useful without requiring an autonomous agent product.

Separate enquiry from commitment

Submitting an enquiry requests a conversation; it is not acceptance of a contract or approval for payment. Label buttons with the actual effect and show an accessible confirmation after success. Preserve validation errors and retry options. Avoid hidden preselected marketing consent and make privacy information available before the form is submitted. Do not interpret a page visit as authority to contact or transact.

Add interfaces only when real

An API catalogue or MCP card should describe working, scoped capabilities. Never publish a fictional OAuth server or action endpoint for an audit score. If an authenticated integration exists, protect tokens, restrict access to the right site and require confirmation for consequential actions. Public information endpoints and private customer workspaces have different access boundaries.

Test failure and handoff

Use keyboard navigation, narrow screens and realistic missing-input cases. Check how the page reports unavailable appointments, failed submission or expired authorisation. An agent should be able to hand the decision back to the person with context rather than silently retry a transaction. “Agentic SEO” is an emerging label, not a standard ranking certification.

Continue with the related decision

• AEO learning paths
• The seven-stage customer journey: where AI search helps and what comes next
• Discuss the scope of AEO and AI-search work