advisorsai-check

Does a public page expose the declared machine-readable basics that answer engines can use?

One command. No API key, no account, no signup.

One click inside your agent (GitHub strips cursor:// links — use the live buttons): https://advisorsai.ai/see.html

npx skills add manmohm/advisorsai-check

/plugin marketplace add manmohm/advisorsai-check

Share a measured page (SSL Labs pattern): https://advisorsai.ai/see.html?url=https://example.com

# one-off, nothing left behind (needs pipx: python -m pip install pipx)
pipx run --spec git+https://github.com/manmohm/advisorsai-check advisorsai-check example.com

# or install once
pipx install git+https://github.com/manmohm/advisorsai-check
advisorsai-check example.com

git clone https://github.com/manmohm/advisorsai-check.git
cd advisorsai-check
python -m pip install .
advisorsai-check example.com

Use it from Claude or any MCP client — no install

The same bounded public-page check is available as one remote, read-only MCP tool. You provide one public URL; it returns at most one actionable finding with an HMAC evidence receipt that the operator can re-verify.

• Add Advisors AI Store Readiness to Claude

• Official MCP Registry name: ai.advisorsai/store-readiness

• Streamable HTTP endpoint: https://advisorsai.ai/store-readiness-mcp

• Inspect the MCP implementation and its operator boundary

The remote tool does not write to the submitted site, store the result, share it, infer answer-engine rankings, or predict sales. It is new: there are no customer case studies yet.

advisorsai-check  https://example.com

  ok   page title: «Example — industrial valves since 1994»
  ok   meta description, 153 chars
  MISS no Organization/LocalBusiness structured data: nothing states what this
       business IS in a form machines read
  MISS no Service/Product structured data: what you sell is only in prose
  ok   one <h1>: «Valves that survive the plant floor»
  ok   canonical link with an HTTP(S) href is present
  ok   declared robots policy allows search crawlers on /
       (training/model-use crawlers blocked by policy: GPTBot, Google-Extended)
  note no /llms.txt (informational only: it is not a web standard)
  ok   valid urlset sitemap with 24 locations

  64% of the declared public-page basics checked in this run are in place.

Related MCP server: aeo-mcp

Run the public-page basics MCP locally

From a source checkout, install the optional MCP dependency with Python 3.10+:

python -m pip install ".[mcp]"
advisorsai-public-page-mcp
# equivalent: python -m advisorsai_check.mcp_stdio

The process speaks MCP over stdin/stdout and waits for a client. It exposes advisors_public_page_basics_check(url), which calls the same bounded public advisorsai_check.run library used by the CLI. Its English report contains url, success, score, coverage_percent, unchecked_count, signals and errors. Each signal contains key, ok, detail, weight and evidence. An incomplete run keeps the library's score: null; coverage is reported separately. As in the CLI, returned text has terminal control channels removed and whitespace normalized; it remains untrusted data. The returned normalized URL retains its query string, so submit a URL that is safe to show in the MCP client.

This portable mode needs no account, API key or HMAC material. It fetches only the supplied public origin through the existing checker's network safeguards. It does not proxy to Advisors AI, issue an HMAC receipt, or provide the official remote tool's one-finding capsule or Arabic/English language option. The official remote receipt-producing mode above remains available separately. Neither mode measures answer-engine rankings, mentions or sales.

For a local MCP client:

{
  "mcpServers": {
    "advisorsai-public-page-basics": {
      "command": "advisorsai-public-page-mcp"
    }
  }
}

Use it as an Agent Skill

npx skills add manmohm/advisorsai-check

Coding agents that follow the Agent Skills format can load SKILL.md at the repo root (GitHub topic ai-skill) or skills/ai-site-readiness/SKILL.md: they teach the agent to run this check, read only observed evidence, and never claim rankings.

Embed the live checker on another README:

[![AI page readiness](https://advisorsai.ai/badge/ai-readiness.svg)](https://advisorsai.ai/see.html)

What it checks

What it does not check

It does not tell you whether an assistant names you when a buyer asks.

That is not established by a page fetch. It is an observation of live answers, and establishing it requires timestamped captures against real assistants and comparison with peers. This tool therefore never relabels its bounded page checks as a visibility score.

So everything here is a public-page basic: a fixable technical condition, not evidence that an answer engine will cite you. Not a ranking. Not a share. Not a promise.

Install from source

git clone https://github.com/manmohm/advisorsai-check.git
cd advisorsai-check
python -m pip install .
advisorsai-check example.com

The package is not yet published to PyPI. Do not use or advertise a pip install advisorsai-check command until the release exists there. The package requires Python 3.10+, including the command-line checker and optional MCP dependencies. The local stdio adapter above is portable. The published official remote-server source is for inspection of its operator deployment rather than a turn-key receipt-producing self-hosted promise. The package pins html5lib==1.1 and webencodings==0.6.1 so malformed HTML is interpreted with one stable HTML5 tree-construction contract and charset labels follow the WHATWG web-encoding registry. The HTML parser and fact extractor run in a byte-capped, time-capped child process; there is no hand-written token or nesting estimator. Home pages must be served as text/html: application/xhtml+xml is deliberately rejected because this tool does not claim an XML parsing contract. Responses must also be an uncompressed identity representation; unsupported Content-Encoding values are rejected instead of being parsed as HTML bytes.

Use it in CI

GitHub Action (Ubuntu runner):

- uses: manmohm/advisorsai-check@main
  with:
    url: https://example.com
    min-score: "80"

--json gives you the full result. Exit code 0 means every checker returned; 1 means at least one check was unavailable or failed internally; and 2 means the submitted address was unusable. A partial run publishes score: null, reports weighted coverage_percent, and is never a successful process result:

set -o pipefail
advisorsai-check example.com --json | jq '.score'

- name: Public-page machine basics
  run: |
    git clone --depth 1 https://github.com/manmohm/advisorsai-check.git /tmp/advisorsai-check
    python -m pip install /tmp/advisorsai-check
    result=$(advisorsai-check "$SITE" --json) || {
      status=$?
      echo "$result"
      exit "$status"
    }
    score=$(printf '%s' "$result" | jq -er '.score')
    echo "Declared public-page basics: $score%"
    [ "$score" -ge 80 ] || { echo "::warning::below 80%"; }

As a library

from advisorsai_check import run

report = run("example.com")
if not report.successful:
    raise RuntimeError(report.errors)
print(report.score)
for signal in report.signals:
    if signal.ok is False:
        print("fix:", signal.detail)

report.successful is false if any signal is unchecked or any checker raised an internal error. In that state report.score is None; use report.coverage_percent to describe how much of the weighted check set ran, never as a substitute score. The report vocabulary is closed: every run carries exactly one receipt for home, title, description, structured_business, structured_offering, h1, canonical, robots, llms_txt, and sitemap, with weights fixed by the library. jsonld_valid is the sole optional diagnostic and may occur at most once. Missing stages, unknown keys, duplicates, empty receipts, or changed weights make the report unsuccessful; they can never yield a score or 100% coverage. The home receipt also pins the six page-derived rows: a checked home failure requires checked failures for title, description, structured business, structured offering, H1, and canonical; an unavailable home requires all six to remain unavailable. The optional jsonld_valid row is accepted only as a negative diagnostic (false) after a full home representation was parsed. It can lower a score but can never act as a positive bonus. Although home and llms_txt are zero-weight stage receipts, leaving either unchecked adds a missing-coverage unit, so a partial run cannot display 100% coverage. signal.ok is True, False, or None. None means the check could not run — a DNS, connection, TLS, or timeout failure, for example. Such failures are reported as unchecked (ok: null), never as evidence that the site failed. They make the whole run incomplete: report.successful is false, report.score is None, and coverage_percent shows only the weighted share that actually ran. Status 0 is reserved for this transport/unavailable state. An HTTP 4xx or 5xx response is different: the site did answer, so it produces a checked result according to that signal's contract before the checker considers the error representation's body or Content-Type. Thus a home-page 4xx/5xx is a checked failure even when the body is empty or mislabeled; robots.txt keeps RFC 9309's distinct 4xx/5xx semantics. A page or robots representation is full only at exact HTTP 200: informational/accepted or partial/delta statuses such as 202, 206, and 226 are rejected. Redirects to another origin are not followed.

The first site we pointed it at was our own

It identified a crawler-policy mismatch between the site's intended robots rules and the rules served at the edge. The lesson is narrower than a ranking claim: inspect what public crawlers actually receive, not only the file in the repository.

We had not noticed. The tool found it in one run, which is the entire argument for running it on yours.

Who made this

Advisors AI measures timestamped answer-engine outputs for a business, and builds systems the client then owns outright. This tool is the free, honest part: the part you can verify yourself, offline, with the source in front of you.

The paid part is the part this tool refuses to guess at — what assistants actually say about you, with the captures to prove it, and what to change.

MIT licensed. Issues and pull requests welcome.