What is agent readability?

Agent readability is how well a website can be discovered, parsed, understood, and acted upon by AI agents — the machine-facing counterpart to human accessibility. It covers discovery files like llms.txt, machine-readable structured data, skill files like AGENTS.md, and protocol manifests for MCP and A2A.

What's the difference between llms.txt and AGENTS.md?

They serve different agents. llms.txt is a curated index that helps retrieval and answer-engine agents find a site's most useful content. AGENTS.md is a skill file that tells coding agents how to build, test, and work inside a codebase. Many projects ship both.

Do I need both an MCP server card and an A2A agent card?

Only if you expose those capabilities. Publish an MCP server card if your site offers tools or resources for AI clients to call; publish an A2A agent card if you run an agent that other agents discover and delegate to. A content site that just wants to be read and cited needs neither.

What is a markdown mirror and why does it matter?

A markdown mirror is a clean-Markdown copy of an HTML page, served at / .md or via content negotiation. It matters because AI extractors parse Markdown far more reliably than HTML cluttered with navigation, scripts, and styling — and several agent-readability checks look for one.

How is agent readability different from SEO?

Traditional SEO optimizes for human click-through from search rankings. Agent readability optimizes for machines that read, cite, and act — emphasizing clean structure, machine-readable manifests, structured data, and content negotiation. They overlap, but agent readability targets answer engines and autonomous agents rather than the blue-link results page.

Agent readability glossary: llms.txt, MCP, A2A & more

This glossary defines the core vocabulary of agent readability: the files, protocols, and concepts that determine whether AI agents can discover, parse, and act on a website. Each definition is deliberately concise.

The canonical references for these terms include llmstxt.org, the Model Context Protocol, the A2A protocol, and Schema.org.

What are the core agent-readability concepts?

Agent readability: The degree to which a website can be discovered, parsed, understood, and acted upon by AI agents — the machine-facing counterpart to human accessibility.
AI agent: An autonomous or semi-autonomous AI system that reads web content and calls APIs to complete tasks on a user's behalf, rather than rendering pages for a person to view.
Answer engine: An AI system such as ChatGPT, Perplexity, Claude, or Google AI Overviews that answers a query directly by synthesizing and citing sources instead of returning a list of links. Optimizing for it is called answer-engine optimization (AEO) or generative-engine optimization (GEO).
Content negotiation: An HTTP mechanism in which a client's Accept header tells the server which representation to return — for example sending Accept: text/markdown to receive a page's Markdown mirror instead of HTML.
Structured data (JSON-LD): Machine-readable metadata embedded in a page using Schema.org vocabulary in JSON-LD format, letting agents and search engines extract entities, relationships, and facts without parsing prose.
Vercel Agent Readability Spec: Vercel's specification defining the discovery files, structured data, and content conventions that make a site readable by AI agents. Agent Ready scores sites against it.
.well-known: A standardized URI prefix, defined by RFC 8615, under which a site publishes machine-readable metadata at predictable paths — for example /.well-known/mcp.json or /.well-known/agent-card.json.

What are the agent discovery and indexing files?

llms.txt: A Markdown file at a site's root that gives AI agents a curated index of its most useful pages and resources, with a short summary and categorized links. Defined by the llmstxt.org specification.
llms-full.txt: A companion to llms.txt that concatenates a site's primary content as a single Markdown document, so large-context agents can ingest everything in one request instead of crawling page by page.
AGENTS.md: A Markdown skill file, conventionally at a repository or site root, that tells coding agents how to build, test, and work within a project. A vendor-neutral convention that supersedes tool-specific files like CLAUDE.md and .cursorrules.
robots.txt: A plain-text file at a site's root that tells crawlers which paths they may fetch, using the User-agent and Allow/Disallow rules of the Robots Exclusion Protocol (RFC 9309). For agent readiness it is where a site separates AI training crawlers from search and user-triggered agents — opting out of model training without losing AI citations.
sitemap.md: A Markdown counterpart to sitemap.xml that lists a site's pages as a readable linked outline rather than raw XML, so agents can crawl the document tree without an XML parser.
Markdown mirror: A plain-Markdown version of an HTML page, served at /<page>.md or returned via content negotiation, so AI extractors can read clean text without parsing markup.

What are the agent protocols and cards?

Model Context Protocol (MCP): An open protocol that lets AI applications connect to external tools, resources, and prompts over a standard JSON-RPC interface. An MCP server exposes capabilities that any MCP-compatible client, such as Claude or ChatGPT, can call.
MCP server card: A JSON manifest at /.well-known/mcp.json that advertises an MCP server's name, description, version, and (optionally) transports so agents can discover and connect to it. Specified in SEP-2127, which superseded the earlier SEP-1649 proposal.
Agent2Agent Protocol (A2A): An open protocol for agent-to-agent communication and capability discovery, letting independent agents find one another and delegate tasks.
A2A agent card: A JSON manifest at /.well-known/agent-card.json that describes an agent's identity, capabilities, and skills under the A2A protocol.
agents.json: A manifest, defined by Wildcard as an OpenAPI extension, that maps a REST API's endpoints to agent-callable actions so agents can invoke them reliably.
agent-permissions.json: A manifest, defined by the LAS-WG at v1.0.0, declaring which actions AI agents may and may not take on a site — verb-and-selector rules plus RFC 2119 guidelines, served at /.well-known/agent-permissions.json.
NLWeb: An open Microsoft protocol that gives a website a natural-language /ask endpoint returning Schema.org-typed results. Every NLWeb instance is also an MCP server.
API catalog: A linkset, standardised by RFC 9727 and served at /.well-known/api-catalog, that lists a site's API description documents (OpenAPI, MCP, A2A) as RFC 9264 link relations so an agent can discover every API in one fetch.
WebMCP: A W3C-incubated browser API (navigator.modelContext) that lets a web page register tools an in-page AI agent can call directly, instead of screen-scraping the DOM. Imperative JavaScript with no static discovery file.
Agent-driven UI (A2UI): Google's declarative format in which an agent returns a JSON component tree that the host renders as interactive UI rather than plain text. Delivered over A2A/AG-UI; the declarative counterpart to MCP Apps' HTML widgets.
Agent-to-Human (A2H): Twilio's open protocol for looping humans into agent work — INFORM, COLLECT, AUTHORIZE, ESCALATE, and RESULT intents delivered over SMS, WhatsApp, push, email, or voice, with signed envelopes and a /.well-known/a2h gateway discovery document.
Agent-User Interaction Protocol (AG-UI): CopilotKit's open, lightweight, event-based protocol that standardises how AI agents connect to user-facing applications, streaming typed events — text, tool calls, shared-state deltas, and human-in-the-loop interrupts — back over SSE or WebSockets.
Web Bot Auth: The IETF effort that lets automated agents cryptographically prove their identity to the websites they visit, using HTTP Message Signatures (RFC 9421): the operator publishes public keys at /.well-known/http-message-signatures-directory and signs each outbound request, so origins can verify a real agent instead of trusting a spoofable user-agent string.

What are the agentic-commerce protocols?

x402: A payment protocol built on the HTTP 402 Payment Required status code, letting agents pay for an API call or resource through a machine-to-machine payment handshake.
Machine Payments Protocol (MPP): An IETF 'Payment' HTTP authentication scheme (from Stripe and Tempo) for agent micropayments — a 402 challenge, a signed payment credential, and a receipt. An alternative payment rail to x402.
Agentic Commerce Protocol (ACP): A protocol for agent-surface checkout — completing a purchase from within an AI agent's interface on the buyer's behalf.
Universal Commerce Protocol (UCP): A protocol focused on merchant interoperability for agent-driven commerce, advertised at /.well-known/ucp.
Agent Payments Protocol (AP2): A protocol for delegated payment authorization, letting a user grant an agent scoped authority to pay on their behalf.

Frequently asked questions

What is agent readability?: Agent readability is how well a website can be discovered, parsed, understood, and acted upon by AI agents — the machine-facing counterpart to human accessibility. It covers discovery files like llms.txt, machine-readable structured data, skill files like AGENTS.md, and protocol manifests for MCP and A2A.
What's the difference between llms.txt and AGENTS.md?: They serve different agents. llms.txt is a curated index that helps retrieval and answer-engine agents find a site's most useful content. AGENTS.md is a skill file that tells coding agents how to build, test, and work inside a codebase. Many projects ship both.
Do I need both an MCP server card and an A2A agent card?: Only if you expose those capabilities. Publish an MCP server card if your site offers tools or resources for AI clients to call; publish an A2A agent card if you run an agent that other agents discover and delegate to. A content site that just wants to be read and cited needs neither.
What is a markdown mirror and why does it matter?: A markdown mirror is a clean-Markdown copy of an HTML page, served at /<page>.md or via content negotiation. It matters because AI extractors parse Markdown far more reliably than HTML cluttered with navigation, scripts, and styling — and several agent-readability checks look for one.
How is agent readability different from SEO?: Traditional SEO optimizes for human click-through from search rankings. Agent readability optimizes for machines that read, cite, and act — emphasizing clean structure, machine-readable manifests, structured data, and content negotiation. They overlap, but agent readability targets answer engines and autonomous agents rather than the blue-link results page.