# Action endpoints Action endpoints Also called agent actions or agentic actions. **In short:** List each action users can take (book, quote, apply, contact) with its URL, and add a line saying when an agent should offer it. - Category: Agent Readiness · weight 10 of 100 within the category - Agent Readiness is worth 30 in technical, 25 in generic and commerce, 20 in regulated (points of 100, by scoring profile) - Audit ID A6 (as shown on reports) · rubric v2.1 ## What it checks Whether the file tells agents how users can book, quote, apply, buy or get in touch. ## Why it matters When a user is ready to act, an agent should be able to hand them the right action — and know when it’s appropriate to offer it. Answer engine optimization ends at the answer; listing your actions is what turns that answer into a booking, a quote or a call. ## When it applies Applies when the file offers an action. Agent-callable: a state-changing API operation (POST, PUT, PATCH or DELETE), a declared MCP endpoint, or WebMCP tools. An MCP server that only searches documentation or is marked read-only takes no action, so it doesn’t count here (it still counts for Machine surface). User-facing: a link titled as an action (“Book an appointment”, “Request a demo”, “Open an account”, “Contact us”) or to an action page such as /book, /signup or /contact, or an action verb with a URL or phone number in prose. A contact link alone is enough. Otherwise it’s Not applicable; it no longer depends on documenting an API. An audit that doesn’t apply, or that we couldn’t verify, is removed from your score and its weight shared among the Agent Readiness audits that do apply. It is never scored as a zero. ## How it’s scored 0.5 for listing an action, always, once the check applies · + 0.5 for guidance on when to offer it: a prose line, or a heading with text under it, that conditions an action on the user (“Use X when the user wants to …”, “If a user wants to book …”) or directs them to one (“direct users to complete bookings on …”). A link description doesn’t count as guidance, because it describes a page rather than instructing the agent. An API or MCP server is not needed for full marks. > **Measurement.** Read from the file’s own lines by fixed patterns, with no model. Link titles are matched whole and short, so “Get started” or “Buy vs. Build” documentation pages don’t count as actions. Nothing is fetched: a linked form or endpoint isn’t opened, and an MCP server is recorded as declared, never called. In our research corpus the check applies to 103 of 187 files, and 10 earn full marks; the mean is 0.549. ## How to fix 1. List each action with its URL (or endpoint) and what the user needs to complete it. 2. Say when the agent should offer it, in a line of its own (“Use … when the user wants to …”), not only in a link description. 3. Include a contact route for anything the actions don’t cover. ## Example ``` ## Actions - [Request a quote](https://acme.com/quote): Quote form for volume pricing - [Book a demo](https://acme.com/demo): Thirty-minute call with a specialist - [Contact support](https://acme.com/contact): For anything the actions above do not cover Use Request a quote when the user asks about pricing above 1M transactions a month. If a user wants to see the product, direct them to book a demo. ``` ## Frequently asked questions ### What should an action endpoints section tell an agent? Each thing a user can do, such as book, get a quote, apply, buy or contact you, with its URL or endpoint, what the user needs to complete it, and when the agent should (and shouldn’t) offer it. Add a contact route for anything the actions don’t cover. Sources: [LLMSRank — Action endpoints audit](https://llmsrank.dev/docs/audits/action-endpoints.md) ### How is the action endpoints check measured today? From the file itself, with fixed patterns and no model: 0.5 for listing an action, plus 0.5 for a prose line, or a heading with text under it, that says when to offer it (“Use X when the user wants to …”, “If a user wants to book …”, or directing users to an action). A link description doesn’t count as guidance. An API or MCP server isn’t needed for full marks. Sources: [LLMSRank — Action endpoints audit](https://llmsrank.dev/docs/audits/action-endpoints.md) · [LLMSRank — How scoring works](https://llmsrank.dev/docs/method.md) ### When does the action endpoints check apply to my site? When the file offers an action. Agent-callable actions are a state-changing API operation (POST, PUT, PATCH, DELETE), a declared MCP endpoint or WebMCP tools; an MCP server that only searches documentation or is marked read-only takes no action, so it doesn’t count. User-facing ones are a link titled as an action (“Book an appointment”, “Request a demo”, “Contact us”) or to a page such as /book, /signup or /contact, or an action verb with a URL or phone number in prose; a contact link alone counts. Otherwise it is Not applicable. It no longer shares the [Machine surface](https://llmsrank.dev/docs/audits/machine-surface.md) gate, so a site with no API can still be scored. Sources: [LLMSRank — Action endpoints audit](https://llmsrank.dev/docs/audits/action-endpoints.md) · [LLMSRank — How scoring works](https://llmsrank.dev/docs/method.md) ### Can an MCP server or WebMCP tool be an action endpoint? Yes. MCP lets agents call tools that take actions on a user’s behalf, and WebMCP lets a page expose actions such as a booking form as tools; list them next to your plain URLs. Tell the agent when to offer each one, and when a human should take over instead. See our [WebMCP and MCP starter kit](https://llmsrank.dev/tools/webmcp) for exposing tools. Sources: [Model Context Protocol — What is MCP?](https://modelcontextprotocol.io/) · [WebMCP — Draft Community Group Report](https://webmachinelearning.github.io/webmcp/) · [LLMSRank — WebMCP and MCP starter kit](https://llmsrank.dev/tools/webmcp) · [LLMSRank — Action endpoints audit](https://llmsrank.dev/docs/audits/action-endpoints.md) ### In our research, how often did action endpoints apply? For more than half of files. In our 187-site corpus (captured 2026-10-10, scored with rubric v2.1) the check applied to 103 files and was Not applicable to 84. 93 scored 0.5 (an action listed, no guidance on when to offer it) and 10 scored 1.0; the mean is 0.549. 64 files list an action as a link and 15 describe one in prose, but only 14 say when an agent should offer it — which is the half of the check almost everyone misses. Sources: [LLMSRank — How scoring works](https://llmsrank.dev/docs/method.md) · [LLMSRank — Action endpoints audit](https://llmsrank.dev/docs/audits/action-endpoints.md) See also: [all audits](https://llmsrank.dev/docs/audits.md) · [how scoring works](https://llmsrank.dev/docs/method.md)