CrawlCheck

Coverage of CrawlCheck's public dataset: how many distinct domains have been measured and how they split by platform, rendering, size, language and kind, with the strata that are still under-sampled.

Where each crawler's path into a domain dies, as an observed data flow: edge decision, robots.txt as a file, robots.txt rules, then the page, JSON-LD, sitemap, llms.txt and entitymap.json it reached. Give agent (e.g. ClaudeBot, GPTBot, googlebot) for one identity's path and a one-line answer; omit it for every identity plus the stores and breaks. Uses the latest record on file, or scans first when there is none (fresh=true forces a scan). The answer-engine output is drawn but never measured.

Draft an llms.txt from the domain's own homepage and up to 25 declared pages, using their titles and descriptions. A draft for the owner to cut down, not a publication.

Why CrawlCheck decided a finding: the rule and its revision, the exact fetches it compared (identity, status, bytes, sha256), when the condition held on the site, and a Merkle proof that the decision is inside the signed manifest. Give id (a d1: decision or f1: finding id from a report) or domain for its top finding. Other findings, and lookups by code, need a licence for that domain (send it in x-crawlcheck-key).

A starter entitymap.json built from the domain's latest scan record: name, phone, coordinates and declared service areas as entities with SERVES relations. Fields the page never stated are marked TODO, never guessed. Needs a prior scan.

A fix plan for MACHINE_CHAIN_HEAVY: the files a crawler reads before the page and the 404s absent agent files return, then measured reductions (llms.txt to an index, a sitemap index, robots.txt comments and duplicate groups, minified JSON, short 404s) with the chain size after each and the step at which the detector stops firing.

A fix plan for PAGE_IS_MOSTLY_CODE: the homepage's measured bytes by category, the largest blocks with what to do with each, ordered steps with the text ratio after each, and the step at which the detector would stop firing - or, when moving code cannot clear it, how much markup to cut or text to add. Give domain (fetched now) or id (a stored scan).

The domain's served robots.txt, corrected: * group rules copied into named groups that lacked them (shadowing), a Sitemap line added when missing, a minimal replacement when the served file was HTML. Every change is listed at the top; nothing else is touched.

A fix plan for STALE_CACHE_SERVED: the homepage fetched now, its Age, the cache layers that name themselves in the headers in purge order (innermost first), the HTML lifetime each response declares, a fresh-read comparison, and whether a purge clears the finding and keeps it cleared.

What a domain declares agents can do (OpenAPI operations, MCP tools, A2A skills, API catalog entries), each classed by safety (read_only_public up to financial), and which ones CrawlCheck actually called and confirmed - plus every mismatch between declaration and behaviour. Only read-only public actions are ever called; everything else says why it was not. From the latest scan on file.

The unified machine record for one observation: subject, grade, findings (the top one in full without a licence), capabilities with verification, and the evidence roots, manifest and verifier links that let anyone check it offline. Give id (report, d1:, f1: or manifest sha256) or domain for the latest.

Which remote MCP servers the official MCP Registry lists on a domain, and what each endpoint actually answered when CrawlCheck sent the MCP handshake (initialized, auth_required, unreachable, tool count, tool-list drift). Use before selecting an MCP tool from that domain.

Call BEFORE reading, citing, connecting to or transacting with a domain. Returns a signed decision: allow, warn, require_confirmation, block or unsupported, with every step that led there and the resolve answer it was made from. Pass your crawler token as agent so robots.txt is checked for you. A policy can only make the decision stricter. Do not proceed on block; ask the user on require_confirmation.

The figures CrawlCheck quotes about itself, read from the source: sites_measured, domains_in_corpus, crawler_visits (a rolling window), sections, sections_scored, findings_published, guides_published, with an at timestamp.

Whether a domain is in CrawlCheck's registry and what it serves: llms.txt, agents.md, a media kit at /.well-known/media-kit.json, a reciprocity-tested entity graph, an AI access policy that names crawlers, and agent-callable surfaces. With no domain, returns the shelf counts across every measured site. A row exists because a fetch produced it; nothing here is self-reported and no payment moves a shelf.

Call before fetching from, citing or acting on a domain. Returns one signed answer: crawler policy as declared (robots.txt per crawler), what each crawler identity was actually served, machine files, declared capabilities with their safety/policy class and whether any was verified, entity reciprocity, and open findings. Each section is declared, observed or not_measured with its time; there is no overall score. An unknown domain returns not_measured (never a guess) and is queued for measurement.

Resolve every named answer engine, search index and training crawler against a domain's robots.txt the way a crawler does: most-specific group only, longest match, allow wins a tie. Shows when a Disallow under * does not apply to an agent with its own group.

Scan a public domain as several crawler identities and return the graded record: grade, AI-visibility score, per-section scores, findings with evidence, and whether the origin refused CrawlCheck. One scan takes 5-15 seconds.

Verified-crawler traffic observed at crawlcheck.io itself: which named agents arrived, how many claims were confirmed against operator ranges, how many were forged.

Given raw web-server access-log lines, decide for each line that names a crawler (GPTBot, ClaudeBot, Googlebot, PerplexityBot...) whether the source IP falls inside that operator's published ranges. Returns verified, spoofed, or unverifiable when the operator publishes no ranges.

Every upload on a YouTube channel, read 50 at a time, sorted by views. The upload list is free; the per-check tally needs a Watch licence and is not offered here.

One YouTube video against the video anchor model: metadata, chapters, captions and storyboard where observed, twelve checks each naming what it reads.

The site half of video: homepage plus up to 60 sitemap pages read for YouTube embeds, facades and page-builder widgets, VideoObject nodes and their required properties, and - with a handle - how many of the channel's videos the site carries.