Blog AI and Agents · Agent Ready · Part 2
We ran our own website through Agent Ready. It scored 39.
An AI assistant sent to book a call with us would have been turned away at the door. Here is what was wrong, what we changed in one afternoon, and the one point we chose not to chase.
We build Agent Ready, a free scan that scores whether an AI agent can find a business, understand it, act on its website and finish a booking. So the fair test was to point it at the preview of our own new website first. It came back 39 out of 100. Grade F. The label on the report was blunt: invisible to agents.
That is a strange result for a site that looks finished. The pages load fast, the copy is clear, every product has a page. A person would have had no trouble. An AI assistant asked to "book a call with Sneeze It" would have failed at almost every step.
Friday morning
39 out of 100. Grade F. Reach 44, Comprehend 44, Act 6. Nine checks failing.
Friday evening
97 out of 100. Grade A. Reach 100, Comprehend 92, Act 97. One small warning left.
What the scan found
The report listed nine open findings. Four of them mattered more than the rest.
- We had told every AI agent to stay out
Our robots.txt disallowed everything, which is normal for a preview. But it also turned away ChatGPT-User, Claude-User and Perplexity-User: the assistants a real customer sends to do a task for them. That one check was worth 10 points.
- Our phone number did not exist
The address, phone and email sat in a footer that JavaScript drew after the page loaded. A person saw them. An agent reading the page saw nothing to call or email.
- Nothing on the page could be filled in or called
No form, no machine readable list of what the business does, and no tools an agent could use. Act, the pillar worth the most, scored 6 out of 100.
- No facts in a form a machine trusts
No structured data, no sitemap, no llms.txt. An agent had to guess what we sell from the design.
Source: Agent Ready scan of one.sneeze.it, October 9, 2026
What we changed
The first pass was plumbing, and it took about an hour. We let the live AI agents in. We published a sitemap and an llms.txt briefing. We added one block of structured data with our real address, phone and email, copied from our own footer rather than retyped. We put the phone and email in the page itself, not in a script. And we turned the "try the free scan" button into a real form with a label, so an agent can type a website into it.
That took the score from 39 to 89.
Source: Agent Ready rescans after each change, October 9, 2026
The last stretch was the interesting part. We registered tools on the page with WebMCP, the new browser standard that lets a site hand an AI agent a short list of things it can actually do. Not a chatbot. A menu of verbs. Ours now lets an agent list our products, get our contact details, check David's real open times, book a walkthrough, send a proposal request, read our Creative Benchmarks, search this blog and run an Agent Ready scan on any website.
A tool that returns a plausible guess is worse than no tool, because an agent will act on it.
Every one of those tools returns real data or says plainly that it could not. The times come from the live calendar. The benchmarks are the published ones. If the calendar will not take a booking directly, the tool says so and hands back a booking link with the details filled in. We did not register a single tool that invents an answer.
The point we did not take
The report still shows one warning, worth one point: our booking runs through Calendly, a third party. We could move booking onto our own pages to collect it. We will not until the booking tool has run against the live calendar for a while. One point is not worth a booking flow nobody trusts.
