Key Takeaways6
- Rankings and readability do not show whether an AI agent completes a booking.
- The AI Agent Readiness audit scores six dimensions from 0 to 100.
- Blocking live fetchers such as ChatGPT-User hurts agent readiness.
- Captchas, forced logins and phone-only booking are the common stall points.
- Most small sites score near zero on programmatic access, so early points are cheap.
- I re-run the audit every 90 days because agent standards move monthly.
Your site ranks well. Your copy reads clearly. A customer tells an AI assistant to book an appointment with you. The agent opens your homepage, follows the booking link, and stops at a captcha.
Search rankings do not measure this failure. A readability check does not measure this failure either. I built a second audit to measure the gap.
Two audits, two questions
My website audit scores six categories from 0 to 100: SEO, AI optimization, performance, compatibility, measurability and accessibility. A site profile sets the weights. A local service business weighs SEO at 25 and AI optimization at 15. A publisher weighs both at 25.
The AI optimization category asks one question. Does a language model read your site and recommend you? I test with five prompts across up to four models: Claude, ChatGPT, Perplexity and Gemini. I count mentions and citations separately.
A mention means the model knows your brand.
A citation means the model retrieved your site.
Each result points to a different fix. Mentions come from long-term content and outside coverage. Citations come from crawlable pages, schema and fresh content.
Recommending is not acting. The AI Agent Readiness audit asks the next question.
The agent readiness question
If a customer points an AI assistant at your site and says "book me an appointment," how far does the agent get?
This question is the whole audit. Booking-driven businesses gain the most: studios, clinics, opticians and salons. Any site with a transaction applies too, such as "buy this" or "request a quote."
Six dimensions
Each dimension scores 0 to 100 in capability points. Here are the largest point values in each.
Bookability. Does the agent reach the primary action? An online scheduler earns 45 points. A booking route on the site earns 25. A web path other than phone-only earns 20. A link in navigation earns 10.
Action schema. Are the actions machine-readable? Structured data, business identity, service schema, a declared action and a reserve or schedule action type earn 20 points each.
Conversational AI. A chat widget earns 40 points. An AI-powered widget earns 35 more. A booking flow to hand off to earns 25.
Programmatic access. An MCP endpoint earns 40 points. A public API earns 30. A plugin manifest earns 20. An agents.json or NLWeb file earns 10.
Agent access. An llms.txt file earns 35 points. Links to booking pages inside the file earn 15. Allowing live AI fetchers in robots.txt earns 25.
Completion friction. No captcha on the path earns 30 points. No forced login earns 25. A web action the agent completes earns 25. No phone gate earns 20.
The overall score is a weighted average. For a local service business, bookability carries 30 percent and friction carries 20. Schema, chat and access carry 15 each. Programmatic access carries 5.
Scores add up from capability points. When you ship a fix, your score moves by the points you shipped.
Indexers and fetchers
Two kinds of bots read robots.txt. Indexers such as GPTBot, ClaudeBot and CCBot crawl in bulk. Live fetchers such as ChatGPT-User, Claude-User and Perplexity-User act when a person asks a question or gives a task.
Blocking the fetchers hurts agent readiness. The audit records the posture for each bot as allowed, blocked, partial or unspecified.
What a script misses
The runner scores what fires on a headless page load, plus the well-known endpoints: llms.txt, MCP, plugin manifests, OpenAPI and agents.json. The runner does not finish a booking. Three checks stay manual.
Chatbot booking. A detected AI chatbot proves nothing about completion. I attempt a real booking through the widget.
Friction on the path. A captcha on a side flow does not block booking. I walk the booking flow to confirm.
Embedded forms. The crawl cannot read a Google Form or a Typeform. I check for forced sign-in and captchas.
The live walk gives the strongest evidence in the report. I open your site as the agent would and note the exact step where I stall.
Check your own site this week
Open your booking page in a private window. Count the steps from homepage to confirmed appointment. A phone-only path fails.
Look for a captcha or a login wall before the action.
Open yoursite.com/robots.txt. Search for ChatGPT-User, Claude-User and Perplexity-User. A blanket Disallow under any of the three blocks the agent.
Open yoursite.com/llms.txt. A 404 means you have no file. Write one and link your booking page.
Ask your developer whether your pages declare a potentialAction with a Reserve or Schedule type.
What to expect
I expect most small sites to score near zero on programmatic access. Few publish an MCP endpoint or a public API. The gap is wide, so the first points are cheap to earn. I will not publish an average score until I have a larger sample.
A high score shows an agent has a path. A high score does not show agents use the path. The audit measures capability, not bookings.
Agent standards move monthly. I re-run the audit every 90 days.
How the two audits fit together
You get the AI Agent Readiness audit as a standalone report. You also get a one-page summary panel inside the website audit report, placed after the roadmap. Both reports share one scoring discipline: the prose findings must explain the score.
Want to know where an agent stalls on your site? Send me your URL through the contact page.
Where would an agent stall on your site?