Is your website ready for AI agents?

A noindex directive hides the page from search engines, and from the agents that rely on them to find their sources.

Fix. Remove the noindex directive from public pages, in the meta robots tag as well as the X-Robots-Tag header.

The list of your public pages. An agent looking for a specific page finds it without crawling the whole site.

Fix. Publish a sitemap.xml listing your public pages and declare it in robots.txt with a Sitemap line.

An agent choosing between ten tools picks the one that says what it is for. The test looks for that guidance in llms.txt or an agents.md.

Fix. Add a When to use section to llms.txt, or publish an agents.md, saying plainly when to recommend your service.

If an agent has to search for your documentation, most give up. Only applies to sites with an API or technical documentation.

Fix. Add a link to your documentation in the homepage menu or footer.

A Markdown summary of the site, at the root or on the documentation, pointing models to your key pages.

Fix. Publish an llms.txt file at the site root, in Markdown, with a title, a summary and links to your key pages. Hikoo's llms.txt generator drafts it in one click.

The llmstxt.org format expects an H1 title, a quoted summary, H2 sections and Markdown links.

Fix. Structure llms.txt per llmstxt.org, with an H1 title, a quoted summary, H2 sections and links.

A broken link in llms.txt sends the agent nowhere.

Fix. Check that every llms.txt link answers 200 and update the file when a page moves.

lastmod dates tell agents which pages changed and are worth reading again.

Fix. Give each page its update date with lastmod in the sitemap.

An MCP server announced in llms.txt or under /.well-known is found by agents without searching.

Fix. Announce your MCP server in llms.txt and in /.well-known/mcp.json or a server card.

Naming AI agents in robots.txt shows their access is intended, not left to the generic rule.

Fix. Name AI agents in robots.txt, for example OAI-SearchBot, Claude-User or PerplexityBot, with the rule you want for each.

A Content-Signal line states whether your content may be used for search, AI answers or training.

Fix. Declare the allowed uses of your content with a Content-Signal line in robots.txt, for search, ai-input and ai-train.

An HTTP Link header points to your agent resources from the first response, without reading the page.

Fix. Add an HTTP Link header pointing to your agent resources, Markdown version, llms.txt or API description.

The /.well-known/ard.json file lists in one place what you offer agents, MCP servers, APIs and skills.

Fix. Publish /.well-known/ard.json to list what you offer agents, MCP servers, APIs and skills.

The page is requested as ChatGPT-User, OAI-SearchBot, Claude-User and PerplexityBot, the agents browsing for your customers. A firewall or a robots.txt rule that stops them makes the site invisible.

Fix. Review your firewall, bot management and robots.txt. ChatGPT, Claude and Perplexity agents must receive the page, not a verification screen or a refusal.

Agents read raw HTML and rarely run JavaScript. The test measures the text delivered, the H1 heading and the share of text in the HTML without scripts.

Fix. Render the main content on the server, with SSR or static generation, with an H1 heading. Aim for at least 5% text in the HTML without scripts by trimming the markup.

An agent without JavaScript gets stuck on a page that redirects with JavaScript or meta refresh, and gets lost when a redirect goes through another domain.

Fix. Serve the site over HTTPS, redirect http to https on the server, and replace meta refresh or JavaScript redirects with HTTP 301 redirects on your own domain.

The test opens several pages linked from the homepage. A page behind a login is invisible to an agent.

Fix. Keep the pages describing your offer, prices and documentation public. Keep sign in for the customer area only.

A missing URL must answer 404. Full credit also requires a Markdown message when the agent asks for text/markdown, so it knows where to go next.

Fix. Answer 404 to unknown URLs, never 200. When the request asks for text/markdown, return a short Markdown message with a link to the sitemap or llms.txt.

An agent asking for text/markdown on the homepage receives a clean version of the page, without menus or scripts, and spends fewer tokens.

Fix. Serve a Markdown version of the homepage and your pages when the Accept header asks for text/markdown, with Content-Type text/markdown and Vary Accept. Keep HTML for browsers.

An agent loads the text of each page into its context window. Beyond 25,000 tokens, it truncates and loses the end.

Fix. Split overly long pages, such as terms or large FAQs, into shorter pages linked together.

A link alternate tag of type text/markdown announces the Markdown version of the page.

Fix. Add a link rel alternate tag of type text/markdown in the head, pointing to the Markdown version of the page.

schema.org JSON-LD states plainly what the page describes, product, article, company or FAQ.

Fix. Add schema.org JSON-LD to the HTML, Organization, WebSite, Product or Article depending on the page.

Title, description, language, canonical URL, og:image and og:type sum up the page before it is read and prevent wrong citations.

Fix. Fill in a title, a meta description, the lang attribute, a canonical URL and the Open Graph og:image and og:type tags.

A complete Organization block ties the site to the right entity and gives agents a way to check who you are.

Fix. Describe the company with an Organization block giving its name, URL, logo, official profiles in sameAs, a contactPoint and the address.

About, Contact and Privacy pages let an agent check who runs the site before recommending it.

Fix. Link your About, Contact and Privacy or Legal pages from the homepage.

An agent comparing options for a customer drops the ones whose price it cannot find. Only applies to sites that sell.

Fix. Make your prices readable by agents, with a pricing page linked from the homepage, Offer JSON-LD or a pricing.md file.

sameAs links to Wikidata, LinkedIn or Crunchbase tie your brand to an entity models already know.

Fix. List at least two official profiles in the Organization sameAs, for example Wikidata and LinkedIn.

Types such as FAQPage, Product, Service or Review hand agents ready made answers.

Fix. Add the types that match the page, for example Service, FAQPage, Product or AggregateRating.

A pricing.md file gives your plans and prices in Markdown, with no page to interpret.

Fix. Publish /pricing.md with your plans, their prices and what they include, and link it from llms.txt.

A single H1, headings without skipped levels and main and nav landmarks give the agent the page outline.

Fix. Keep one H1 per page, heading levels without gaps and main and nav landmarks.

Hidden text giving orders to agents is treated as prompt injection and loses their trust.

Fix. Remove any instruction aimed at AI from attributes and invisible text. Say what you do plainly, in visible content.

For a site offering an API, the OpenAPI specification is how an agent integrates it without reading the docs. It is looked for at the usual locations and in the documentation.

Fix. Publish your OpenAPI specification at a stable URL, for example /openapi.json, and link it from the documentation and llms.txt.

When a call fails, the agent needs a structured JSON error, not an HTML page. The test calls a missing path of the API.

Fix. Make your API answer in JSON on every error, 404 included, with a readable code and message.

An agent must be able to get access in a standard way. The test looks for the OAuth metadata of your API or MCP server.

Fix. Publish your authorization server metadata at /.well-known/oauth-authorization-server (RFC 8414) on the API or MCP server host.

Named scopes let you give an agent only the access it needs. The test looks for them in the OpenAPI specification and RFC 9728 metadata.

Fix. Declare named OAuth scopes, for example read and write, in the specification securitySchemes or in scopes_supported of the protected resource metadata.

The test calls the API announced by the specification, or api followed by your domain, and checks that it answers.

Fix. Expose your API at a stable address, announced in the OpenAPI specification, that answers in JSON even without authentication.

Each operation needs an operationId and a description so an agent knows which one to call.

Fix. Give each operation in the specification a unique operationId and a one sentence summary.

An agent turns an operation into a tool when it has an ID, typed parameters and a typed response.

Fix. Type the parameters, body and responses of each operation with JSON schemas in the specification.

A response described by a schema lets the agent read the result without guessing its shape.

Fix. Describe the 2xx response of each operation with a JSON schema in the specification.

Errors described by a shared schema let the agent understand the failure and fix its request.

Fix. Define an error schema in components.schemas and reference it in the 4xx and 5xx responses.

Agents avoid an API that can change without warning. The test looks for a version in the URLs and a deprecation policy.

Fix. Version the API, for example with /v1, and flag what is going away with deprecated in the specification or the Sunset and Deprecation headers.

Rate limit headers let an agent slow down on its own instead of being blocked.

Fix. Return the RateLimit or X-RateLimit headers on API responses, and Retry-After on 429s.

An MCP server lets assistants like Claude or ChatGPT use your service directly. The test queries it without an account.

Fix. Keep your MCP server authentication and check with an authorized client that its tools list correctly.

A command line tool lets coding agents use your product without writing an integration.

Fix. Publish an official command line tool on npm, PyPI or Homebrew, and mention it in llms.txt.

The RFC 9727 /.well-known/api-catalog file points to your APIs and their specifications.

Fix. Publish /.well-known/api-catalog as a linkset, with a service-desc link to each OpenAPI specification.

WebMCP exposes the page actions, search, booking or cart, as tools an agent calls in the browser.

Fix. Expose the page's key actions as WebMCP tools with navigator.modelContext, so an agent calls them instead of guessing the interface.

A skills index offers coding agents ready made instructions to integrate your product.

Fix. Publish /.well-known/agent-skills/index.json to offer coding agents ready made instructions.

If an agent represents your service, its A2A card lets other agents find it and talk to it.

Fix. If an agent represents your service, publish its A2A card at /.well-known/agent-card.json.

A portal to sign up, get a key and manage an integration without emailing anyone.

Fix. Publish developer documentation with self serve sign up and API keys.

A protected MCP server must tell the agent where to authenticate, with a WWW-Authenticate header pointing to its OAuth metadata.

Fix. Return a 401 with WWW-Authenticate Bearer and resource_metadata pointing to /.well-known/oauth-protected-resource.

RFC 9728 metadata tells the agent which authorization server to use for your API or MCP server.

Fix. Publish /.well-known/oauth-protected-resource on the API or MCP server host, with resource, authorization_servers and scopes_supported.

PKCE protects the OAuth sign in of an agent that cannot keep a secret.

Fix. Support PKCE with the S256 method and announce it in code_challenge_methods_supported.

A clickable div does not exist for an agent reading the accessibility tree. Only real buttons, links and fields are usable.

Fix. Replace clickable div and span elements with button, a and input tags.

Browsing agents click from the accessibility tree. A button without a name is a button they do not use.

Fix. Give every link and button a text or aria-label, and every image an alt attribute.

An agent fills in a form by reading each field's label. Without a label, it guesses.

Fix. Tie a label to each field with the for attribute, or give it an aria-label.

Cursor pagination lets an agent walk a list without skipping or repeating items.

Fix. Paginate API lists with a cursor, for example cursor or starting_after, described in the specification.