Falling behind.

My website wikipedia.org has no WebMCP tools registered. WebMCP (the W3C draft at https://webmachinelearning.github.io/webmcp/, shipping behind a Chrome origin trial and supported in the ChatGPT desktop browser) lets a page expose typed tools that AI agents can call instead of guessing at the UI.

Task: identify the 2-3 core actions a visitor performs on this site (e.g. search, check availability, get pricing, submit an inquiry) and register each as a WebMCP tool on page load:

document.modelContext.registerTool({
  name: 'search_products',            // snake_case, specific
  description: 'Search the catalog; returns names, prices and URLs.', // written FOR the model: what it does, when to call it, what it returns
  inputSchema: { type: 'object', properties: { query: { type: 'string' } }, required: ['query'] },
  async execute({ query }) { /* call the site's existing search endpoint and return JSON */ }
});

Guard for environments without the API (if (!('modelContext' in document)) use the @mcp-b/global polyfill or return early). Wire execute() to existing endpoints — do not invent new backend functionality. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org has HTML forms but none are annotated as declarative WebMCP tools. The WebMCP draft defines a zero-JavaScript path: annotate a form with toolname/tooldescription attributes and agents can invoke it as a structured tool.

Task: find the site's most important forms (search, contact, booking, newsletter) and add declarative tool attributes, e.g.:

<form toolname="contact_sales" tooldescription="Send a message to the sales team; expects name, email and message." action="/contact" method="post">

Keep names snake_case and descriptions written for an AI model (what it does + what it expects). Do not change form behavior for human users. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org registers WebMCP tools but serves no Chrome origin-trial token, so the modelContext API does not exist for real Chrome users (Chrome 149-156 gates WebMCP behind an origin trial).

Task: register the site's origin for the WebMCP origin trial at https://developer.chrome.com/origintrials, then serve the token on every page that registers tools — either <meta http-equiv="origin-trial" content="TOKEN"> in the <head>, or an Origin-Trial HTTP response header. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org serves no MCP server card. Agents probe /.well-known/mcp/server-card.json to discover whether a site offers a backend MCP server.

Task: if the site has (or should advertise) an MCP endpoint, publish a valid JSON server card at /.well-known/mcp/server-card.json describing it (name, description, endpoint URL, transport). If there is no MCP server, this check is optional — skip it rather than publishing a card that points nowhere. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org's pages are missing <link rel="canonical"> tags, so agents and indexes cannot resolve one authoritative URL per page.

Task: add a canonical link tag to every page's <head> pointing at the page's clean absolute URL (no query strings or tracking params). In most frameworks this belongs in the shared layout using the current route. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org has no /llms.txt. llms.txt is a curated markdown map that tells LLMs what a site is and where its key content lives (spec: https://llmstxt.org).

Task: create /llms.txt as static markdown: an H1 with the site name, a one-paragraph summary, then sections linking the site's most important pages with one-line descriptions each. Keep it under ~200 lines, factual, and free of marketing fluff. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org serves /llms.txt but no /llms-full.txt. llms-full.txt is the companion file with full flattened content for LLM ingestion.

Task: generate /llms-full.txt containing the full markdown content of the site's key pages (docs, guides, product pages) concatenated under clear H1/H2 headings, ideally regenerated at build time so it never goes stale. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org has no valid Schema.org JSON-LD structured data. Agents lean on JSON-LD to understand what a site is, sells, and offers without parsing prose.

Task: add <script type="application/ld+json"> blocks to the site's layout/pages covering its core entities — Organization (or LocalBusiness) sitewide, plus the type matching each page (Product with offers, Service, Article, FAQPage…). Validate every block parses as JSON and passes https://validator.schema.org. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org has a missing or trivial <title> tag.

Task: give every page a specific, descriptive title of roughly 50-60 characters ("What it is — Site name" pattern), unique per page. When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.

My website wikipedia.org sends no Permissions-Policy header. WebMCP tool exposure is governed by this policy (the "tools" directive, default self), and declaring one signals deliberate configuration.

Task: add a Permissions-Policy response header at the server/CDN layer declaring the powerful features the site actually uses and disabling the rest, e.g.: geolocation=(), microphone=(), camera=(). When you are done, verify the fix by re-scanning at https://webmcp-ai.dev/score/wikipedia.org (or POST the URL to https://webmcp-ai.dev/api/scan) and confirm the check now passes.