Strong technical baseline

Server-side render your homepage so AI crawlers see meaningful content without JavaScript. Ensure an H1 and 500+ chars of text in raw HTML.

Allowlist known AI agent User-Agents (ChatGPT-User, ClaudeBot, Google-Extended, DeepSeekBot) in your WAF or bot-detection rules.

Stop blocking user-triggered agents in robots.txt. Remove any 'Disallow: /' that applies to ChatGPT-User, Claude-User or Perplexity-User - including a blanket 'User-agent: * / Disallow: /', which covers them by default. These agents fetch a page only because a person just asked about you, so a block turns away your highest-intent traffic.

Replace meta-refresh and JavaScript-only redirects with real HTTP 301/302 redirects. Non-JS agents never execute location.href or wait for a meta refresh - they see only the stub page. Verify with curl -sI <url> - you should see a Location header, not a 200 with a near-empty body.

Serve your content pages without a login wall. Agents cannot complete auth flows while browsing - a 401/403 or a login-form page is invisible content. Keep public documentation public; if some content must stay gated, publish an ungated summary so agents can still represent it.

Publish an OpenAPI (Swagger) specification at /openapi.json or /api/openapi.yaml. This is how agents understand your API surface automatically.

On the responses that serve text/markdown via Accept negotiation, add Accept to the Vary header (Vary: Accept, Accept-Encoding). Without it, CDNs can serve the cached HTML variant to an agent asking for markdown (or vice versa), depending on which variant landed in cache first.

Verify that major agent User-Agents can reach the homepage. If robots.txt, your WAF, or bot rules block them, remove or narrow the blocking rule. Add an allow rule only when your security setup denies them by default.

Return structured JSON error responses with error codes, messages, and resolution hints. Agents can't parse HTML error pages.

If your MCP server advertises the resources capability in its initialize handshake, make sure resources/list returns at least one resource. If you don't intend to expose resources, omit the capability - the check returns na with no penalty for tool-only servers. Quality of the resources you do return is scored separately by mcp-resource-quality.

Return a real HTTP 404 (or 410) status for nonexistent paths - never a 200 with your app shell, which makes agents believe every path exists. For full credit, give the 404 response a short markdown body pointing agents at your sitemap, llms.txt, or docs index. Verify with curl -s -o /dev/null -w "%{http_code}" https://yourdomain.com/some-path-that-does-not-exist - it must print 404.

Make your developer resources (API docs, OpenAPI spec, auth docs, webhooks, MCP server) discoverable by name. Publish them at predictable URLs, list them in llms.txt, and include your product name in page titles and headings so search engines surface them for name-based queries.

Make sure a clean search for your brand name returns your own domain in the top results. If it does not, your brand may be too generic, conflict with a more established term, or not yet indexed. Strengthen brand-name search by claiming consistent NAP across listings, earning press mentions that link to the canonical domain, and avoiding redirect chains that mask the apex domain in search results.

Add explicit AI crawler directives in robots.txt. Allow the crawlers that feed answer engines ('User-agent: GPTBot' / 'Allow: /', same for ClaudeBot, PerplexityBot, OAI-SearchBot), and restrict training-only crawlers (CCBot, ByteSpider) with 'Disallow: /'. Content Signals ('Content-Signal: search=yes, ai-train=no') earns the same credit.

Add a valid XML sitemap at /sitemap.xml listing all indexable URLs. Include lastmod dates and keep it under 50MB.

Reduce markup overhead so readable text is at least 5% of your HTML. Strip unused inline scripts/styles, server-render content instead of shipping large JSON hydration blobs, and keep wrapper nesting shallow.

Add JSON-LD structured data to your homepage using the identity type that matches your site - SoftwareApplication for products, Organization or LocalBusiness for companies, Person for personal sites, Article for blogs - with name, description, url, and type-appropriate fields (offers, sameAs, author) so AI can parse your identity programmatically.

Publish API documentation at a discoverable URL (/docs, /api, /developers). Include authentication, endpoints, and example requests.

Tell agents when to reach for you: add a 'when to use this' section to your llms.txt (or a dedicated agent-instructions file) that names your best-fit use cases and how an agent should call you. Be specific about the jobs you are right for - generic marketing copy does not read as guidance.

Add all four signals to your homepage: , , , and . Agents use these for entity resolution and attribution.

Add Organization JSON-LD that includes both contactPoint (with email/phone and contactType) and address (PostalAddress). This lets AI verify your business legitimacy and answer contact queries.

Publish real /about, /contact, and /privacy pages with at least 500 characters of content each. These are the pages AI agents check to verify your business is legitimate before recommending you.

Keep each page's extracted text under ~100K characters (~25K tokens) so it fits an agent's context window without truncation. Split oversized reference pages into focused per-topic documents and link them from an index. Check a page with curl -s <url> | wc -c and remember agents read the extracted text, not the raw HTML.

Close every fenced code block (``` or ~~~) in your served markdown. CommonMark treats everything after an unclosed fence as code, so an agent parsing the document silently loses the rest of it. Count fence lines per file - the total must be even.

Create a developer portal at /developers with API keys, documentation, quickstart guides, and a sandbox environment.

Expose a public REST or GraphQL API. AI agents need programmatic access - not just a web UI - to integrate with your product.

Build an MCP (Model Context Protocol) server exposing your API as tools. Use Streamable HTTP transport for full score. This lets Claude, ChatGPT, and other AI agents call your product natively.

Return standard rate-limit headers on your API responses (the RFC RateLimit headers, plus Retry-After on a 429) so agents can self-throttle in real time, and document the conventions alongside your API.

Document your error responses in your OpenAPI spec: give 4xx and 5xx responses a typed error schema (or use RFC 9457 application/problem+json). A consistent error object with a machine-readable code and a human-readable message lets agents handle failures without guessing.

Declare a versioning policy agents can rely on: version your API (in the URL path or a version header) and publish how you signal deprecation (a Sunset/Deprecation header or a documented timeline). Agents avoid integrating against a surface that can change without warning.

Publish an official CLI tool on npm, PyPI, or Homebrew. A CLI lets agents and developers script interactions with your product without building API integrations from scratch.

Define typed JSON response schemas for every endpoint in your OpenAPI spec. Agents rely on these to know what fields they will get back; missing or partial schemas force trial-and-error.

Make your API spec self-describing: a unique operationId and a description on every operation, typed parameters, and response schemas. For GraphQL, a fully typed schema with a documented cost or rate limit reads best.

Ensure API endpoints have unique operation IDs, typed schemas, and descriptions compatible with LLM function-calling formats.

Ensure every resource returned by resources/list reads cleanly via resources/read: declare a valid mimeType, return non-empty content, and make sure any URIs in the content resolve. Broken or empty resources break agent UX silently.

Publish a JavaScript/TypeScript SDK package on npm so developers can integrate your API programmatically. In package.json set repository to your source repo and homepage to your product domain - these links are how agents confirm the package is your official SDK rather than a third-party tool with a similar name.

Serve your MCP server at /.well-known/mcp, publish a server-card.json at /.well-known/mcp/server-card.json, or reference it in llms.txt so agents can discover it automatically without manual URL input.

Create an llms.txt file at your domain root (/llms.txt) - the AI equivalent of robots.txt. Write at least 100 characters of real content: what your product is, what it does, and links to your key docs. Then verify it with curl https://yourdomain.com/llms.txt - you should see your text, not HTML. If your app returns its homepage for every URL (common with single-page apps), add a static file route so the raw text is served. A placeholder with just a heading earns no credit.

Format your llms.txt as a navigation index: start with a markdown heading, include markdown links to deeper resources, and keep it under 30,000 characters. If you have more to say, move long-form content into /llms-full.txt or per-section files (e.g. /docs/llms.txt, /api/llms.txt) and link to them from the main index.

Add sameAs links in your JSON-LD structured data pointing to your Wikipedia page, Wikidata entry, GitHub org, and social profiles. This helps AI disambiguate your brand from similarly named entities.

Make every link your llms.txt declares resolve to real content. Verify each one with curl -L <url> - you should see the linked document, not your homepage. If your app returns the homepage shell for unknown paths (common with single-page apps), a 200 status is not proof: check the body. Fix or remove any dead link; agents that follow the index treat a broken link as a dead end.

Pick one: (a) return Content-Type: text/markdown on GET when the request sends Accept: text/markdown, or (b) publish a static /llms.md, /auth.md, or /agents.md file at your root with real markdown content. Option (b) is usually a single static file. This is the cold-discovery path for agents that land at your homepage from web search without reading llms.txt first.

Add detailed descriptions (>= 20 chars) to every MCP tool. Agents use these to decide which tool to call - vague descriptions lead to wrong tool selection.

Set server name, version, and instructions in your MCP server's initialize response. Instructions help agents understand your server's purpose and constraints.

Expose 3+ tools via your MCP server's tools/list endpoint. Cover your core API surface - agents need tools for read, write, and search operations.

Use consistent naming conventions (snake_case or camelCase) for all MCP tools. Names should be descriptive (>= 4 chars) and not generic (avoid 'run', 'get', 'do').

Protect your MCP server with OAuth 2.0 authentication. Publish authorization server metadata at /.well-known/oauth-authorization-server for automatic agent auth flows.

Return structured JSON-RPC errors (with code and message) when agents call invalid tools or pass bad arguments. Don't crash or return empty responses.

Upgrade your MCP server from legacy SSE to Streamable HTTP transport. Streamable HTTP is the current standard and supports bidirectional communication.

Support an idempotency key on your write operations and declare it where agents can read it: an Idempotency-Key header parameter on your POST/PUT/PATCH operations in your OpenAPI spec for REST, or a client-supplied id argument on your GraphQL mutations. Agents retry on network failures, and without this a retry can double-charge or duplicate a record.

Publish official SDK packages across multiple language ecosystems (npm, PyPI, Go modules, RubyGems). Auto-generate them from your OpenAPI spec using tools like openapi-generator. For each package set the project URL or homepage to your product domain (package.json repository/homepage, PyPI Home-Page or project_urls, RubyGems homepage_uri) - this is how agents verify the package is your official SDK.

Add behavioral annotations (readOnlyHint, destructiveHint) to your MCP tools. Agents use these to avoid destructive actions without user confirmation.

Publish a server card at /.well-known/mcp/server-card.json describing your MCP server. Required fields: name, description, version, serverUrl, tools[]. This lets agents preview your server before opening a transport connection.

Offer a batch endpoint that accepts an array of operations in one request, documented in your spec, so an agent acting on many items can do it in bulk instead of looping one call at a time.

Add MCP Apps support to your MCP server using @modelcontextprotocol/ext-apps. Expose ui:// resources and add _meta.ui.resourceUri to tools so agents can render interactive UIs directly in conversation.

Support Agent-to-UI rendering via MCP Apps (ui:// resources), OpenAI Apps SDK, or generative UI patterns that let agents render interactive UIs in conversation.