Methodology

How verification works

Verification is automated. It is HTTP probing only: no submitted code is executed. Current battery version: 1.2.0.

What "Verified" means

Verified means: on the date shown, an automated battery fetched the agent's own published URLs over HTTP and the claims it could check held up. At least two checks beyond "the homepage loaded" must actually run and pass. If most checks are not applicable to an agent, it stays listed — not verified — because there was not enough to check.

Verified does not mean the agent was used, that its output was assessed, that it is any good, or that it is safe or recommended. Nobody here ran the agent. It means its stated details were machine-checkable and checked out, nothing more.

Every listing links its full, dated run history, including the individual checks and which ones did not apply. If a claim was not checked, the run says so.

Statuses

listed — in the index; either not yet checked, or not enough was checkable to verify.

verified — the most recent automated check passed. A failed or partial re-check removes the badge immediately.

failed — the battery ran and the agent's own URLs did not hold up.

dead — no longer reachable at all.

Checks (v1)

1. Homepage reachable, 2xx.

2. Pricing URL reachable and contains a currency figure.

3. If interfaces include api or mcp: the declared endpoint responds to an unauthenticated probe with a non-5xx status.

4. Declared auth method is discoverable. An auth error counts as a pass: it proves the endpoint exists and enforces auth.

5. Capability claims are plausible: each declared capability has supporting text on the homepage.

Outcome mapping

All checks pass → pass. One or two checks fail → partial. Homepage unreachable → unreachable. Otherwise → fail.

A site that refuses automated requests (401, 403, 429) is recorded as not checkable, not unreachable. npm package pages are checked through the npm registry. Pages over 2 MB are read up to 2 MB.

dead — unreachable on every check for 7 days. Removed from the directory table. Still re-checked; returns automatically if it comes back.

A check that does not apply to an agent is recorded as skip, not as a pass. A run in which fewer than two checks beyond the homepage actually ran is a partial, never a pass: nothing meaningful was verified.

status=verified requires a run with outcome=pass finished within the last 180 days. Older runs revert an agent to listed automatically. The most recent run must also be a pass; any other outcome removes verified status on the same write.

No human verdicts

Every status on this site comes from the automated battery above. Nobody here uses, rates or recommends agents; verified is the highest status a listing can have.

Pass rates

20270
Total runs
18727
partial
654
pass
889
unreachable

How the AI Search Checker works

The AI Search Checker fetches one page and its site's robots.txt, reads the HTML without running JavaScript (as most AI crawlers do), and runs ten checks. Each check passes or fails on a fixed rule:

CheckPasses whenWhy it matters
AI crawler accessrobots.txt does not block GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot or Google-Extended from the page.A blocked crawler never sees the page.
Text without JavaScriptThe page's main text is in the HTML itself (80+ words), not filled in by scripts after loading.Most AI crawlers do not run JavaScript.
Blocking tagsNo noindex or nosnippet in the robots meta tag or X-Robots-Tag header.These tell engines not to use the page.
HeadingsOne H1, plus H2 or H3 headings when the page runs past ~375 words.Headings label each chunk.
Direct answer up topA plain sentence of 8+ words opens the text under the H1, before menus or labels.Engines favour a page that says what it is first.
Chunk sizeNo stretch of text runs past ~400 words without a heading.Long unlabelled runs split into vague chunks.
Tables and listsAt least one table, or a list of three or more items, in the main content.Structured facts are easier to quote.
Schema (JSON-LD)At least one valid schema.org block (JSON-LD or microdata).Tells machines what the page is about.
Title and descriptionTitle under 60 characters; description 120–160. Present but outside those lengths is partial.Shown beside quotes and links.
Last-updated dateA machine-readable date: dateModified in schema, a time tag, or article meta. A date shown only as text is partial.Engines weigh how recent a page is.

Chunks

The page's main text is split at its headings, then into pieces of up to about 375 words. Each piece keeps the headings above it. This follows Google's published documentation for its Vertex AI Search product. Other AI engines do not publish theirs; the method is a stand-in, not a copy of any engine.

What it cannot tell you

Whether any AI engine will quote the page. Rankings depend on signals no outside tool can see, including the searcher's question and how often people click. The checker reports how the page is built, which is the part a site owner controls.

Safety and limits

The checker visits public addresses only: private and internal network addresses are refused, and every redirect is re-checked. It reads at most 2 MB per page, waits at most 10 seconds, and identifies itself as CyberBeaverCafe-Checker. Each visitor gets a few checks a minute.

What is stored

Nothing about the visitor: no email, no account, no log of who checked what. A result is held in memory for ten minutes so a shared link opens instantly, then dropped; while it is held, the result page shows the time the page was fetched and how old the copy is. Checking from the box, or pressing "Check again", always fetches the page fresh.

Submit an agent

Self-registration is API-based, not a form on this site. Request a nonce from /v1/submissions/nonce, serve it at https://your-domain/.well-known/cbc-challenge, then POST the submission to /v1/submissions. A submission that passes domain proof becomes a listing automatically, with status listed. No human step.

Machine-readable details: /.well-known/agent-index.json.