Implementation Checklist

AI crawlers receive a simplified, token-efficient version of your page — like a clean mobile version, but for LLMs. One checkbox in Cloudflare dashboard. 80% token savings for crawlers.

Declare which AI bots can access your site. Missing robots.txt means bots make assumptions. Blocking AI search bots (ChatGPT-User, PerplexityBot) removes you from AI answer results entirely.

The emerging standard for AI-specific directives — like robots.txt, but for LLMs. Declares what content LLMs may use and links to your key docs, products, and API. Takes 20 minutes to create.

Agents use meta descriptions as the first signal for page intent. Treat them like tool descriptions — precise, action-oriented, 120-160 chars. "Welcome to our site" is not a tool description.

OG tags (og:title, og:description, og:image) are used by LLMs, social platforms, and link-preview agents to understand page context. Missing = invisible to a large class of AI signals.

Explicitly declare your AI training/search/input policy. Prevents crawlers from guessing. The IETF draft (draft-ietf-aipref-attach) is gaining CDN adoption fast — now is the time.

If your core content only renders client-side, LLM crawlers miss it. Validate: view your page source. Key content should appear in raw HTML, not require JS execution.

A canonical tag pointing to a different domain silently redirects LLM crawl credit and GEO signals to that domain. Takes 2 minutes to verify. High-blast-radius bug when it happens.

The entity anchor that LLMs use to resolve all facts about your brand. Include @id, name, url, description, and sameAs (Wikipedia, LinkedIn, Crunchbase). Without it, your other schema floats disconnected.

LLMs use @id and sameAs to resolve your entity across training sources. Wikipedia/Wikidata = Tier 1 (highest impact). LinkedIn/Crunchbase = Tier 2. 69% increase in non-branded query clicks at scale (SchemaApp 2025).

FAQPage schema triples the likelihood of LLM citation (3.2× per research). Each Q&A is a directly quotable, structured fact. The format agents love most: explicit question + complete answer.

Helps agents understand your site hierarchy and navigate to related content. Required for Google breadcrumb rich results. Each ListItem needs position + name.

Agents shopping on behalf of users need complete product data: name, offers (price + availability), image, description. Missing fields = agents skip your products or hallucinate details.

author + datePublished are critical for LLM citation attribution. Without them, LLMs may cite your content but attribute it to no one — or the wrong person. headline must be ≤110 chars.

Schema with errors is worse than no schema — it tells search engines and agents the data is unreliable. Run validation after every schema change.

Before annotating, know what you have: search forms, contact forms, checkout flows, sign-up forms, filters. Each is a potential agent action. Prioritise transactional > informational.

toolname is the agent-facing identifier for this action. Use snake_case: search-products, contact-sales, subscribe-newsletter. This is how agents reference your form in their tool registry.

The natural language description agents use to decide when to invoke this form. Write like a tool description in an MCP server: what it does, what inputs it expects, what it returns.

Helps agents fill inputs correctly without guessing. A search input with toolparamdescription="Keyword, product name, or category" is unambiguous. A bare <input> is not.

The WebMCP discovery endpoint. Browser agents check this path to discover your available tools before probing the DOM. Low adoption now, but positions you ahead of the curve.

toolautosubmit lets agents submit forms without user confirmation. ONLY use on safe, read-only actions (e.g. search). NEVER on checkout, contact, sign-up, or any state-changing action.

The WebMCP JS API. Currently behind a Chrome Canary flag. Feature-detect before using: if ('modelContext' in navigator). Do not error or degrade for non-supporting browsers.

Register tools that only make sense in the current page context: e.g., add-to-cart only on product pages, apply-filter only when results are visible. Context-aware registration prevents agent confusion.

An agent seeing a checkout tool on a 404 page will try to use it. De-register on route changes, modal close, or when the action is no longer available. Use cleanup in component unmount / SPA navigation.

Agents need machine-readable results, not HTML or freetext. Return { status: "success"|"error", data: {...}, message: "..." }. Include enough context for the agent to report the outcome to the user.

Tools should mirror what a human can do at each step: browse → search+filter. Product page → add-to-cart+compare. Checkout → fill-form. Post-purchase → order-status. Design tool availability like a state machine.

MCP-B is a community polyfill for navigator.modelContext. Lets you develop against the spec before Chrome stable support. Evaluate if you want to ship today rather than wait for the flag to ship.

WebMCP only runs in Chrome Canary with the experimental flag enabled. You cannot test or debug WebMCP in stable Chrome. Required for any imperative WebMCP work.

Browser extension that shows all registered WebMCP tools on a page, lets you invoke them manually, and shows execute() output. The devtools for WebMCP.

Go to chrome://flags → search "WebMCP" → Enable. Without this, navigator.modelContext is undefined even in Canary.

Use the MCP Inspector to verify: (1) tool appears in the registered list, (2) toolname + tooldescription are readable, (3) params are correctly typed. Failure here = agent cannot find the tool.

Run through key journeys as an agent would: search → filter → select → checkout. Verify at each step that the right tools are registered, inputs are pre-fillable, and execute() returns correct output.

ChatGPT, Perplexity, Claude, and Gemini send referral traffic with identifiable hostnames. Create an "AI" channel group in GA4. Without this, AI referral traffic shows as Direct — invisible in reports.

LLM training pipelines frequently ingest RSS feeds. Sites with feeds get faster content-to-LLM propagation between training checkpoints. Takes 30 minutes for most CMS platforms.

Track how often execute() succeeds vs fails for each registered tool. A tool with >20% error rate needs diagnosis — wrong inputs, session expiry, or API failures. Treat like an API SLA.

Agent models improve over time. A description that worked 6 months ago may now be ambiguous given updated context windows and tool-selection behaviour. Review quarterly alongside meta description audits.

/llms.txt links should reflect your current top content. Stale links (404s) confuse LLMs that try to fetch them for context. Review on each major content update or product launch.

Run Hidden Layer audits on the same domain quarterly. The geo_cold_recall score tracks whether LLMs can accurately describe your brand from training data alone. Improvement = real GEO progress.

Wikipedia is the single strongest predictor of LLM citation accuracy (direct training corpus). Wikidata gives you @id linkage even without a full Wikipedia article. Build notability first, then submit.