Methodology

What the test measures

The test checks what an AI agent receives when it fetches a product’s public documentation. Each check is a direct HTTP request. The response is the evidence, and the report quotes it.

No AI model reads the docs or scores anything. Anyone can repeat a request and get the same answer from the same server, so every result can be reproduced.

What gets audited

Each product is audited at the documentation URL recorded in its report. The audit fetches that page and the site-wide files that agents use: robots.txt, llms.txt, llms-full.txt, the sitemap, Markdown versions of the page, Model Context Protocol endpoints, and standard discovery files under /.well-known/.

The audit reads only what the server returns. It does not run JavaScript, because most agents and crawlers do not. It sends the user agent docs-auditor/1.0 and no cookies, and it never signs in.

The audits use the SaturnDocs docs auditor. Each report prints the audit date and the scoring version.

The AI Agent Readiness Score

The score is out of 100 points in three groups. Content delivery is worth 50 points, discovery 25, and agent capabilities 25. The percentage becomes a US school letter grade, A+ at 97 and above down to F below 60. Percentages print to one decimal place, rounded down, so the printed number never crosses a grade boundary.

• Content in server HTML (15 points). The main content of the page has at least 250 characters of text in the HTML response, before any JavaScript runs.
• Markdown mirror (15 points). A Markdown version of this page, at a .md address or for an Accept: text/markdown request, earns 10. It earns 5 more when it matches the page: its first heading shares a word with the HTML heading or title, and it has at least a quarter of the page text.
• Page fits one agent fetch (10 points). The Markdown version, or the main content of the server HTML, is at most 50,000 characters for full points and at most 100,000 for half. Agent tools commonly cut off longer responses.
• AI crawler access (10 points). The share of AI search and user-triggered fetch agents that robots.txt allows. Rules for training crawlers do not count.
• llms.txt (15 points). A valid index earns 7. It earns 4 more when it is under 50,000 characters, and 4 more when its sampled .md links return Markdown.
• Sitemap and robots.txt (5 points). robots.txt exists, declares a sitemap, and the sitemap is valid XML.
• Page metadata (5 points). JSON-LD, a canonical link, a meta description, and Open Graph tags in the server HTML.
• MCP server (15 points). A Model Context Protocol server that completes the initialize handshake without sign-in earns 15. A server that asks for sign-in earns 7, because an agent cannot use it for the docs without an account.
• API description (5 points). An OpenAPI document earns 3, and an API catalog at /.well-known/api-catalog earns 2.
• Agent interoperability signals (5 points). One point each, up to 5, for Content Signals, an agent skills index, OAuth discovery, Link headers, a Web Bot Auth directory, an MCP server card, and agent sign-in instructions.

Two extra-credit items earn 1 point each: llms-full.txt and an embedded docs AI assistant. Extra credit can restore lost points, but it cannot raise the score above 100. llms-full.txt is extra credit because most of these files are far larger than one agent request can read. A docs AI assistant is extra credit because it serves human readers, not external agents.

The docs platform is recorded as metadata. It is never scored.

Inconclusive checks

A request that times out or fails at the network level is inconclusive, not failed. An inconclusive check earns no points, and the report labels it NOT VERIFIED.

The audit retries a product up to three times when a check is inconclusive, and it keeps the most complete result.

What to fix first

Every report lists the checks that lost points, ranked by the points a fix would add back. Each item names the change that would earn those points. The fixes are documentation and delivery changes, never product changes.

Audit dates

A score reflects the documentation on its audit date, which every report prints. A new audit replaces the score and prints its own date. To request a new audit after a change, nominate the docs site again.

Limits of the method

The audit checks one page and the site-wide files around it. It does not check every page.

The audit measures delivery: whether an agent can find and read the docs. It does not judge whether the docs answer a developer’s questions correctly. The retired reading test measured that with a panel of AI models.

A site that blocks or slows the auditor’s requests can score lower than it would for a real agent. Inconclusive checks are labeled on every report.

The retired reading test

From August to October 2026, the series used a reading test. Three AI models read each product’s public docs and attempted five first-hour developer jobs, and five agent surface checks added points. The reading test was retired on October 16, 2026. Its reports stay in the reading test archive with their original scores, which are not comparable with the current scores.