Colour Intl · agent readiness audits
See exactly what an AI agent gets from your site.
Eight read-only requests, no JavaScript executed, about fifteen seconds. You get the transcript of what came back, an agent readiness score out of 100, and the one thing standing most in the way — on this page, with no account.
The four seconds you never see
A customer asks an assistant. Your site has to answer without you.
There is no browsing, no scrolling, and nobody looking at your design. The assistant fetches a few files, reads whatever is in the HTML before any JavaScript runs, and decides whether you are one of the answers. Here is that exchange, in full:
Who near me does this, and can I book it for Friday?
Three places publish what they offer and how to book one. I could not read the rest, so I left them out.
- ✕GET /.well-known/agent.json404
- ✕GET / — Accept: text/markdowntext/html
- ✕Schema.org on the homepagenone
- ✕GET /sitemap.xml404
- no description it can quote, no action it can nameskipped
That is the whole job. We measure those requests, rank what is missing by what it costs you, and build the fix.
Nothing broke. Nothing was logged. No bounce, no error, no dip in analytics you could point at — you were simply not one of the three. It is invisible from the inside, which is why it goes unfixed for years, and the only way to see it is to make those same requests yourself and read what comes back. So that is what the free scan does, and what it hands you.
What you get
A readout like this one, about your site, in about fifteen seconds
This is a real result, not a mock-up. Booksy is one of the ten platforms on our published
benchmark, so the number is checkable — and you can re-run this exact scan on
booksy.com yourself.
55/100
C — partially readable, material gaps
- 8 passed
- 5 failed
- 1 warning
- 3 not applicable
Largest gap — 12 points of the 100
A capability manifest is published
none of /.well-known/agent.json, /.well-known/ucp, or /.well-known/ai-plugin.json returned valid JSON
This is the file that tells an agent what your site can do — which actions exist, at which paths, returning what. Without it an agent has to infer your site’s capabilities from guesswork, and it will get them wrong.
Fix. Publish a capability manifest at /.well-known/. If you
sell online, follow the Universal Commerce Protocol shape so agents can read your catalog and
drive checkout without scraping. one to three days
…and four more failures and a warning, each with its own evidence, what it costs you, and the fix — plus the transcript of all eight requests.
Three checks did not apply to Booksy and were left out of the total rather than counted against it. Nothing here needs an account, and nothing is installed. Give an email only if you would rather have the whole thing saved to a link and sent to you — and if you never do, the findings are still yours to hand to your developer.
What the scan reads
Sixteen weighted checks, eight read-only requests, one score out of 100
The audit is unauthenticated and changes nothing. It looks at the six things that decide what a machine finds when it arrives, and a check that cannot apply to your site is left out of the total rather than counted against you:
- Discoveryrobots.txt, a sitemap that is both referenced and real,
llms.txt, and a capability manifest - Policywhether AI use is stated separately from permission to crawl —
Disallowcannot say “index me but do not train on me” - Identitystructured data saying who you are, a canonical URL, Open Graph for the card that gets quoted
- Machine statewhether any page will answer in Markdown or JSON instead of a layout built for eyes
- Renderingwhether the content is in the HTML at all, before a line of JavaScript runs
- AccessCORS on the machine endpoints,
Vary: Accept, HTTPS that actually resolves
You get a score, a grade, and the single largest gap named on the spot. Every finding is yours to hand straight to your developer at no cost — the method, the weights and the limits of each check are set out in the published method, and the long version of this argument walks through what was changed on a real site.
Evidence
One site taken from 32 to 100, and an industry that has not started
haecceity.me is a live production site. It was scored, worked on for a day, and scored again — same scanner, same site, same day. No copy was rewritten and no keywords were added; the change was six pieces of engineering.
Ten of the largest salon booking platforms, scanned August 2026
Median 54 of 100, best 67, worst 31 — and not one of the ten publishes a capability manifest, the file that tells an assistant a booking can be made and how. A separate sweep of nineteen independent salon websites the same month came back at a median of 56. Every score and the method, or the other industries we have measured.
What it costs
The scan is free. What is priced is doing the work and keeping it true.
Everything the audit finds is yours whether you buy anything or not. If you have a developer, hand it over and we are glad to have been useful. There are a dozen free tools that will hand you a number, and you should use them — what none of them do is the work. This is the only one where the same scanner that gave you the number then grades our own fix, in public, against a threshold we do not get to move.
Two checks sit outside the guarantee because they are not ours to fix: whether a page renders without JavaScript, which belongs to your framework, and HTTPS, which belongs to your host. Both are said out loud before any money moves. The full terms, and the six changes in detail.
Who you would be buying from
Sean Sneed, trading as Colour International LLC
Colour Intl is an independent studio making websites legible to the machines that now read them. One person, in San Francisco, who does the work himself. No account manager, no sales team, and no call you have to sit through before seeing a number. That is a fair thing to be cautious about, so here is what you can check rather than take on faith: California entity 202461516261, registered 22 March 2024, independently owned with no outside investment — searchable by name or number at the California Secretary of State, and listed on LinkedIn and GitHub.
We do not promise rankings, traffic, revenue, or a place in any AI answer. Agent-driven attribution is not measurable today and any figure for it is a guess. The claim is narrower and checkable: your site is unreadable to these systems right now, and afterwards it is not. The full company record lists what to verify and where, including the corrections we have published against descriptions of this company held elsewhere.
Practising it
This site publishes everything it sells
Read any page here with Accept: text/markdown
and you get the content without the chrome. The rest is where a machine expects to find it:
The convention for llms.txt is young and the manifest format is younger. We publish both, in the open, so that the studio can be audited on the same terms it audits everybody else.
So — what does an agent get from your site?
Type the address. The same engine that produced the Booksy readout above runs on yours, on the spot, and names the largest gap straight away.