The standard
The AgentReady standard, in plain language
AgentReady is an open standard for the agentic web. It sets out, in three stages, what a site needs so AI agents can find it, read it and act on it.
Stage 1 of 3
Find
Can agents reach you and cite you?
Be reachable and citable. Don't block AI agents or crawlers, and give them one clear, current page to cite.
Agents start by finding your site. If they are blocked, or cannot tell which page is the authority on a topic, they move on to a competitor.
- MUST
Don't block AI agents or crawlers
Block agents in robots.txt, a firewall or bot protection and they cannot reach your content at all. Blocking is one of the two changes that made agents fail outright.
- SHOULD
One canonical, current docs page for key answers
Give each important answer one authoritative page that is kept up to date, so agents cite the right source instead of a stale copy.
- SHOULD
sitemap.xml referenced from robots.txt
A sitemap lists your public pages. Pointing to it from robots.txt helps agents discover everything you want them to read.
- MAY
Optional AI Catalog at /.well-known/ai-catalog.json
A machine-readable catalog of what your business offers. It is an emerging convention, so treat it as a useful extra rather than a requirement.
Stage 2 of 3
Read
Can agents read your answers?
Put the answer where agents read. Answers should be in the first HTML a visitor receives, with honest status codes and structured data.
Most agents read the page as it arrives from the server. If the answer only appears after JavaScript runs, many agents will never see it.
- MUST
Answers in the initial HTML, not injected by JavaScript
The words agents need must be in the HTML the server sends. Content that JavaScript adds later is invisible to many agents. This is the other change that made agents fail outright.
- MUST
Honest HTTP status codes: real 404s, and 429 with Retry-After
A missing page should return a real 404, not a friendly page with a 200 status. When you throttle agents, return 429 with a Retry-After header so they know when to come back.
- SHOULD
A raw-HTML homepage with real links to docs, product and pricing
Your homepage should link plainly to the pages that answer the most common questions. Agents usually start there.
- SHOULD
Docs that answer real user tasks, with fenced code snippets
Write pages around what people actually try to do, and show code in clearly marked blocks that agents can copy without guessing.
- SHOULD
An llms.txt index, linked from the footer and metadata
llms.txt is a short index of your most useful pages for AI systems. Link to it from the footer and page metadata so it can be found.
- SHOULD
JSON-LD structured data
Structured data states facts such as who you are, what you offer and your questions and answers in a form machines can read without guessing.
- MAY
Optional markdown mirrors of your pages
A plain markdown copy of a page, offered alongside the HTML. Agents chose markdown 96% of the time when they were asked for a format, so mirrors are cheap to offer.
- SHOULD
Pricing published as text on a first-party page
Put your prices in plain text on your own site. Answers built from your own pricing page were accurate 60% of the time, against 37% from other sources.
Stage 3 of 3
Act
Can agents act for your customers?
Let agents act, not just read. Offer at least one interface agents can use to complete a task, such as a booking, an order or an inquiry.
Reading tells a customer about you. Acting lets an agent complete the task for them. This stage is needed for products that agents act on, and it is optional for content-only sites.
- MUST
At least one interface agents can act through
An MCP server or a CLI lets an agent complete tasks, such as a booking, an order or an inquiry. This is a must for products that agents act on.
- SHOULD
Agent-completable authentication
Agents should be able to sign in on behalf of a user through API keys or OAuth with discoverable metadata, without a person needing to click through a web form.
- SHOULD
OpenAPI contract, SDKs or CLI, and WebMCP tools in the web UI
A published OpenAPI contract, official SDKs or a CLI, and WebMCP tools in the web interface give agents a clear, tested way to act.
- MAY
Agent payments: x402, MPP, UCP or ACP checkout
Emerging payment protocols let an agent complete a purchase, for example through checkout flows such as Stripe Link. Optional, and still early.
- MAY
An NLWeb /ask endpoint for content-heavy sites
A question-and-answer endpoint that lets agents query a large body of content directly. Useful for content-heavy sites, and optional for most businesses.
How to read the badges
Every requirement carries one level. The levels tell you what to fix first.
- MUST
agents fail without it.
- SHOULD
measurably improves results.
- MAY
emerging conventions.
Access and usage
Being AgentReady does not mean giving up control
Being AgentReady means agents can reach and read your public content. You still decide who may visit, what they may see and how your content may be used.
Most owners we work with block a few paths, such as admin areas, carts or private pages, and keep their public content open. That is a normal and healthy setup.
AI-usage preference signals let you separate who may visit from how content may be used. For example, you can opt out of AI training while still allowing AI answers and search. We help you set these deliberately. [Ben to confirm wording before launch.]
Decide what's off-limits
Block only low-value or private paths, such as admin areas, carts or internal pages. Everything public stays open to agents.
Let verified agents through
Keep bot protection on, and allow the agents you trust to pass through it.
Throttle politely
When traffic is heavy, respond with 429 and a Retry-After header, so agents know when to come back instead of failing silently.
Set your AI-usage preferences
Separate who may visit from how content may be used. For example, opt out of AI training while still allowing AI answers and search. [Ben to confirm wording before launch.]
Where the standard comes from
AgentReady is written by ora.ai and Vercel, is at version 2.0 (October 2026), and is open source under the MIT licence.
The standard is grounded in measurement: 1,033 real agent runs across 25 public product sites such as Stripe, Notion, Datadog and Twilio, from June to August 2026. A controlled fetchability experiment added 190 probes against one site under 19 configurations.
Read the full standard, with the dataset behind it, at the canonical site.
agentready.org