Methodology
How the score is made
A score about a named company has to be checkable. This page says what we measure, what we do not, and how to read a number without over-reading it.
1
What is measured
138 checks of public pages and endpoints, in four layers. Each check earns points. A layer’s score is the share of its points earned. The layers are weighted, and a layer with no applicable checks is left out and its weight is shared among the others.
- Discovery
- 20%
- Can an agent find you, and does your site say what it is?
- Access
- 30%
- Can an agent get in and read your docs and API description?
- Use
- 40%
- Can an agent do the work: call the API, use the MCP server, sign in?
- Payments
- 10%
- Can an agent pay? Counted only if you support a payment path.
Checks that do not apply to a site (for example payment protocols on a site with no payments) are removed from the sum, not scored as failures. So are checks we could not run, such as a page behind a login. Those are called unverified, and the report shows how many there are next to the score. Payments use a best-path rule: only the payment method that scored best counts.
2
The gate
3 checks decide whether an agent can read the site at all: crawler reachability, bot detection, and whether the content is in the HTML without JavaScript. If any of them fails, the score is capped at 59, because nothing else matters much if the agent cannot get in. A page that shows our crawler a challenge is undetermined, not failed, and does not cap the score. The report says which of the two happened.
3
Where each claim comes from
Every number carries one of these labels, and the labels are never mixed in one number.
- Static
- A check of a public page, file or endpoint. This is the whole score today.
- Observed
- What our crawler, MoatlightBot, actually received: allowed, challenged or blocked, signed and unsigned.
- Declared
- What robots.txt says about other AI crawlers. We never send a request as GPTBot, ClaudeBot or any other vendor’s crawler, so for them we only know what the file says, not what happens.
- Simulated
- An agent run through a task, either by hand or by our read-only agents. Shown in its own section, never in the score, and with "not enough data" when there are fewer than 5 runs.
- Real traffic
- Agents seen in a site’s own logs. Not part of the free scan.
4
How deep a check goes
Of the 138 checks, 50 are deep: they fetch, parse and compare, and they cover what matters most for a developer product (gate, content, docs, OpenAPI, MCP, auth). 70 are shallow: they look for a file or a signal and say little about quality. 18 are not built yet and always report “unverified”. The report marks each one.
What a scan never does
- Sign up, log in, submit a form or call an API that changes anything.
- Call a tool on an MCP server. We list tools and read descriptions only.
- Use another vendor’s crawler name.
- Run real coding agents against your product. That is the paid, weekly step, and it is labelled Simulated when it appears.
Our crawler identifies itself as MoatlightBot/1.0 (+https://moatlight.tech/bot) and signs some requests with Web Bot Auth. Details and how to opt out are on the MoatlightBot page.
5
Why our number may differ from another scanner's
Other tools use the same four-layer idea, and ours is not a copy of any of them. We wrote our own checks and wording. Where a check is not built yet we leave it out of the sum rather than guess, and where a deep check reads a site differently from a shallow one, the results differ. Expect the biggest gaps in the Use layer. If we are wrong about your site, tell us at the address in the footer and we will look at the evidence in the report, which shows the request and the response behind each result.
6
Rules and versions
- Rules version
- 1.0.0-r1, published 2026-11-16
- Rules hash (SHA-256)
- a56e272fd66ef2ac812dfd7ad39eeeaee4648728ab3c221f708f74dd882e453b
- Grade bands
- A+ 95–100 Front of the packA 86–94 StrongB 70–85 CapableC 48–69 PatchyD 28–47 Hard goingF 0–27 Shut out
Every report shows the rules version and hash it was scored with. Any change to the weights or the gate changes the version, so an old score can be traced to the rules that made it. A change never rewrites an old report.