Blog Inside Sneeze It · Inside the AI layer · Part 4
How Agent Ready checks a website: 17 checks, two passes and a wall detector
Our free scanner asks whether an AI assistant can actually book or buy on your site. Here is exactly how it decides, and the rules that keep the score honest.
Agent Ready answers one question for a gym, studio or spa owner: if a customer asks an AI assistant to book or buy from you, can the assistant actually get it done on your website? Anyone can run it, free and with no account, at agentready.sneeze.it. We have written about what it found across 98 fitness and wellness websites, and about taking our own site from 39 to 97. This post is about the instrument itself: what it checks, how it looks, and the rules we built in so that a score means something.
Four doors, weighted by what matters
An AI agent sent to your website hits four questions in order. Can it get in? Can it understand what you sell? Can it actually do the thing, whether that is booking a class, requesting a quote or buying a membership? And can it finish without running into a captcha or a login wall? Those are the four pillars, Reach, Comprehend, Act and Complete, and between them they hold 17 scored checks worth 100 points.
The weights are the opinion in the product, so we will state it plainly. Reading your page is table stakes. Doing something on it is the point. Act carries 35 of the 100 points, more than any other pillar, and its heaviest single check, worth 14 points, is whether the site offers WebMCP tools an agent can call.
Source: Agent Ready methodology
| Pillar | What it checks (points) |
|---|---|
| Reach (25) | AI crawlers allowed in robots.txt (10), the site serves a request from an agent (7), a secure and predictable address (4), a sitemap an agent can follow (4) |
| Comprehend (25) | Machine readable business facts (10), content readable without JavaScript (7), an llms.txt briefing for AI (4), the page states what it is (4) |
| Act (35) | WebMCP tools an agent can call (14), forms a machine can fill in (9), a machine readable capability manifest (5), phone and email as real links (4), a booking or checkout the agent can walk through (3) |
| Complete (15) | No captcha blocking the last step (6), public information stays public (3), a second request still works (3), few third party frames in the way (3) |
The robots.txt check looks for 12 AI crawlers by name, including GPTBot, ClaudeBot, PerplexityBot and Google-Extended. The total maps to five grade bands, from A, "Agent ready," at 90 and above, down to F, "Invisible to agents." Sites that sell products online get a fifth pillar, Commerce, with six store checks such as product offer data and a cart that works without JavaScript. It applies only to online stores, so a gym or med spa is never marked down for not having a shopping cart. Every check and point value is public on the methodology page.
Source: Agent Ready methodology and scanner code
Two passes: what a cheap agent sees, and what a real browser sees
The scanner only reads public pages. It never logs in, never submits a form and never writes anything. It identifies itself honestly as an agent, then reads the homepage, robots.txt, llms.txt, the sitemap, any MCP discovery files and a handful of the page's scripts.
That first pass is plain HTTP, which is roughly what a simple AI agent sees. Where it could change the answer, the scanner then runs a second pass in a real Google Chrome with the experimental WebMCP feature switched on.
WebMCP is worth a sentence of explanation, because it is the biggest single check. It is an experimental Chrome feature, from Google and Microsoft, that lets a website hand an AI agent a short menu of named actions, like "book a class," instead of leaving the agent to read the screen and guess where to click. A scanner that only reads code can find signs that a site mentions WebMCP. It cannot tell whether the tools are really registered when the page runs. The live pass can: it opens the page and lists the tools the site actually registers. A confirmed live list scores higher than code that merely mentions WebMCP, and the report says which kind of evidence it used. That live pass has been running in production since October 9.
- 1Plain fetchRead the public pages the way a simple agent would
- 2Wall checkDid a bot wall or a challenge page answer instead of the site?
- 3Real ChromeOpen the page with WebMCP on and list the tools it registers
- 417 checksScore each one, with the evidence shown
- 5ReportGrade, four dials, fixes in order, a WebMCP starter kit
When a wall answers instead of the site
Bot protection is the most common way a site in our industry turns agents away, and it is invisible to the owner, because humans see the page normally. So the scanner treats it carefully. When a wall answers the plain fetch, the scanner retries in the real browser to measure what sits behind it, so the report can still tell you about your content and your booking path. But Reach still scores zero, because a real agent was turned away at the door, and a report that hid that would be flattering you.
There is a harder case. Some protection services do not answer with an error at all. They serve a challenge page, the "checking your browser" screen, with a normal OK status. An early version of the scanner took that page at face value. A well-known massage chain's homepage returned a security challenge page with an OK status, and the scanner scored it as an empty site, a report about a website it never actually saw. In the 98-site research, the scanner flagged 16 sites as walled, and we found two more by hand, which is why that post counts 18. Walls like those two are the reason for the fix.
On October 10 we shipped a fix. The scanner now recognizes challenge pages from the major protection services even when they arrive with an OK status, treats a near-empty page with no real content as a suspected wall and checks it in the real browser, and, if the browser lands on an error page rather than the site, scores nothing behind the wall instead of scoring the error. The report names what it found and why it counted as a wall. That massage chain's homepage is now reported as a wall, not as an empty site. The fix has been live on agentready.sneeze.it since October 10.
Before the October 10 fix
A challenge page served with an OK status is scored as the homepage. The report describes an empty site that nobody actually saw.
After
The challenge page is recognized as a wall. Reach scores zero, the evidence says why, and the rest of the report measures only what the scanner could truly see.
Rules, not a model, write the report
Every finding, plain-language line and fix in an Agent Ready report is fixed text chosen by the checks. No AI model writes it. That was a deliberate choice: a scanner that tells you what is wrong with your site has to say the same thing twice for the same site, and it cannot be allowed to make anything up.
A check we could not run says so. It is never scored as a pass or a fail.
That rule matters more than it sounds. If a pillar has nothing to measure, it drops out of the total instead of quietly passing. When we added the rule, a bare test page with almost nothing on it lost points it had been given for checks that never ran. Those points were generosity, not measurement.
The same principle runs through the starter kit the report writes for the biggest gap, the WebMCP layer. It proposes two or three named actions an agent could call, each tied to something the scan actually saw on your page, and each labelled either "likely exists" or "must build." It will never propose a booking tool that invents your availability, because a tool that makes up an open slot sends a customer to an appointment nobody is expecting. And the scanner passes its own test: Agent Ready publishes its own robots.txt, llms.txt and MCP discovery file, and registers three WebMCP tools of its own, to run a scan, explain a check and describe the scoring method.
