# How the agent-readiness scan works We diagnose public, machine-readable website signals. The result is a versioned technical baseline — never a certification, search ranking, or promise of model behavior. ## What we check The scan runs 18 checks across discovery, structured content, agent-like HTTP access, signals for your kind of business, and experimental surfaces. - `#1` robots.txt availability - `#2` AI crawler policy - `#3` Content-Signal policy - `#4` Sitemap - `#5` Structured data availability - `#6` Business-specific structured data - `#7` Markdown content negotiation - `#8` llms.txt - `#9` MCP server card - `#10` UCP profile - `#11` OAuth discovery metadata - `#12` Server-rendered content - `#13` Agent user-agent access - `#14` Page weight and response time - `#15` Product feed discovery - `#16` Platform fingerprint - `#17` A2A agent card - `#18` Language alternatives ## How scoring works The backend excludes not-applicable checks, removes unavailable weight from the assessed denominator, and stores the resulting score together with coverage and rubric version. > Coverage below 70% produces an incomplete, provisional result. Coverage below > 30% produces no public score, and neither does a home page the site answers with > a bot challenge or a refusal, whatever the coverage: a 401, 403 or 429 status, a > known challenge vendor's page, or an error page saying access is denied. > Benchmarks stay hidden until at least thirty scans of the same segment have > finished with coverage of 70% or more. Coverage is shown as the whole percent > it has reached, rounded down, so 69.8% reads 69% and only a scan that assessed > everything reads 100%. - Invisible: 0–24 - Readable: 25–49 - Callable-ready: 50–69 - Ahead of the market: 70–100 ## What each result means - Pass — the expected public signal was verified. - Partial — a signal is present but incomplete. - Fail — an accessible surface confirmed a negative fact. - Unavailable — we could not reach a verdict; this is not a failure. - Not applicable — the check does not apply to this site or segment. A result other than a pass names one cause, and its evidence lists the fields, sitemaps or codes involved. ## What the scan could not read The scanner reads each response up to a limit of its own: 512 KiB for robots.txt and llms.txt, 1 MiB for a discovery document such as an MCP server card, a UCP profile, an A2A agent card or OAuth metadata, 2 MiB for a page, and 5 MiB for a sitemap. A longer response is cut at its limit, and a result that depends on it names the limit. A sitemap served as a gzip file is unpacked, and the 5 MiB limit applies to what it unpacks to. A gzip sitemap that cannot be unpacked is reported as broken, since an agent cannot read it either. A response that sends more than twice its limit and 64 KiB without decoding to the limit is stopped, and the check that needed it is unavailable. The scan reads at most three sitemap files, an index and the sitemaps it lists included, and a listed sitemap whose address redirected counts as read at the address that was listed. What was read proves what it contains. A heading, a link or a sitemap entry found before the limit counts. A discovery document longer than its limit is too large, since its size is known, and a page longer than 2 MiB fails the page-weight check for the same reason. An llms.txt longer than its limit passes when a heading and a link were read before the limit and is too large otherwise. A robots.txt longer than its limit is partial, since crawlers need read only its first 500 KiB. A sitemap may be larger than 5 MiB, so it is judged by the entries read, and one cut before any entry is not judged. A check that would have to conclude from a cut page that something is missing is unavailable. A 404 or 410 says a document is not there, and so does a web page where a document belongs: an answer that opens with a page's doctype or `` tag after any white space, comments and XML prolog. An empty answer is no document where a JSON document belongs. An empty llms.txt is malformed, an empty sitemap is not a sitemap, and an empty robots.txt passes, since the robots standard reads it as having no rules. Any other 4xx, such as 401 or 403, refuses the document to an agent as it did to the scanner, so the scan takes it as absent and says so: the check fails, or does not apply to the optional llms.txt and A2A card, and the result asks you to allow the document or answer 404. An MCP server card whose address asks for a sign-in fails in words of its own, since an agent has to read the card before it can sign in. No answer at all, a timeout (408), a rate limit (429), a refusal of the scanner's request headers (431), an answer that depends on where the scanner asks from (451) or a server error says nothing about the document, so the check is unavailable and its evidence gives the status. A JSON document that is there but is not a valid JSON object, whatever it opens with, is reported as unparseable. When a check reads several documents, a defect proven by what was read stands, while a pass or a verdict that nothing usable exists needs every document read. One complete MCP server card passes whatever the other location answered. A sitemap index whose listed sitemaps did not all answer is not judged as a whole unless a defect was proven. The evidence names each sitemap that did not answer and, while the scan has read fewer than three files, each that the index lists on another site, which the scan does not read. ## Non-scoring browser observations When enabled, the separate `browser-public-v2.0.0` layer passively renders public pages to observe JavaScript-only content, accessibility/form semantics, runtime/network health, metadata consistency, declared agent/API surfaces, and rendering cost. These findings are shown separately and never change the 18-check `gtm-v1.0.0` score. ## What the scan cannot tell you A regional HTTP measurement can observe latency, output, metadata, and access at scan time. It does not prove worldwide availability, crawler identity enforcement, live stock or appointment correctness, or how any particular model will answer. > “Callable-ready” describes public discovery metadata. It does not mean that > Agentify invoked a tool, authenticated a user, completed a booking, or verified > checkout. ## Checked, observed, and owner-only - **Checked:** public crawler policy, sitemap, substantive initial HTML, JSON-LD structure, and the optional llms.txt surface. - **Observed only:** public protocol declarations, passive browser structure, form semantics, and a bounded sample of visible-versus-structured facts. - **Requires verified ownership:** authentication, tool calls, forms, booking, checkout, payments, rate-limit testing, idempotency, receipts, internal logs, SLOs, and kill switches. ## What we do not do We do not log in, submit forms, click actions, bypass robots or access controls, scan non-standard ports, perform vulnerability probing, make bookings, execute checkout, or change the target site. The optional passive browser layer executes only the public page JavaScript needed to render it and blocks mutating requests.