AgenticWebScore
Scan

Agent readiness checker: See your site the way agents do.

60 checks from a public rubric. Each shows its evidence and its source.

Read a report render.com neon.com cloudflare.com

Free. No account and no email address. A site that blocks every automated client is reported as not measured, not as a low score.

The scanner uses a plain HTTP client and public rubric v0.9.1. Each check lists a shell command that repeats its measurement. Read the methodology

How the score is composed

essential ×3 · recommended ×1

  • Web standards 28 checks 70%
  • Agent conventions 8 checks 30%

Web standards, by dimension

  • Discoverability 37%
  • Content 54%
  • Crawler access 9%

API and agent checks and commerce checks join the web-standards part only when the scan finds that surface on the site. When no agent convention applies, or none was measured, web standards are the whole score.

Agent readiness tools

See every tool

Each checker here runs the full scan, and the report opens on the checks it names. The generator writes robots.txt lines and sends no request.

Start with the guide: How to make your website agent-ready

Then read: What is agent readiness? A plain guide · What is a good agent readiness score? · What is llms.txt? Does Google use it? · How to test an MCP server with curl

Anatomy of a scan

What happens between pressing Scan and reading a number

vercel.com · 75 requests · 23.3s · rubric v0.9.1

  1. 0.43s / the homepage
  2. 0.06s /robots.txt robots
  3. 2.18s /sitemap.xml sitemap
  4. 0.09s /llms.txt llms
  5. 0.15s /.well-known/security.txt .well-known security
  6. 1.25s /security.txt security
  7. 0.09s /about about
  8. 23.3s Verdict recorded — 28 checks scored, L3 Machine-readable Fingerprint c7be46d7b492 · 8 to fix

Read from a stored record, not a mock-up. The full transcript, every response size and every status code, is in the status file for this scan — and the same log streams live while yours runs.

The measured corpus

What the scanner found on the published candidates

We measured 5 of the 5 published candidates under rubric v0.9.1. The highest score is 88.5 (cloudflare.com), and the median is 73.6. 2 of the 4 sites with an API or agent surface reached level 4. 3 checks were not passed by any of them.

Measured
5
of 5 · v0.9.1
Highest
88.5
5 sites · v0.9.1
Median
73.6
5 sites · v0.9.1
At L4
2
of 4 with an API or agent surface

Read from the stored scan records of the 5 published candidates; the newest was measured 2026-11-11. Measured under the current rubric v0.9.1.

Illustration: a grid of dim squares, crossed by a scanning beam that comes to rest, with a ring-shaped iris watching the grid.
Illustration, not data: no square stands for a measured site.

Restraint

The boring decisions, stated plainly.

A block reports “not measured”, never a low score

A site that refuses automated clients is not failing. It is not measured. Scoring it anyway would punish a decision it is entitled to make, so the score is withheld and the reason is published.

Nothing earns points that was not observed

If a probe gets no usable answer, the check is reported as not measured. It leaves the score, the confidence figure falls, and a level gate it holds is not reached.

Cautious about what it fetches

Every redirect hop is resolved and refused if it points at a private or link-local address. At most 3 scans run at once, and each client may start 10 per minute. The scan requests 24 paths on every site, and up to 82 when more checks apply. Beyond them it requests only a capped set of URLs the site's documents name, and 1 random missing path, so it is not a crawl.

Readable without JavaScript, here and in the scan

Every page renders completely with JavaScript disabled, and loads 1 script from another origin: the page-view counter at analytics.marifeting.com. The homepage also loads /hero-glow.js from this site, which only moves a glow with your pointer. The scan runs none of the scripts on your site. It reads what a client without a browser can read, because that is the client it measures for.

AgenticWebScore measures agent readiness: how well a website serves an agent that has no browser. It fetches a public domain from the outside, the way an agent would. It scores what it observed against a versioned rubric of 60 checks in 5 dimensions. Every defect lists what was observed and, where one exists, a command that reproduces it.

Lighthouse runs in a browser. This scanner uses a plain HTTP client, and it gives no score to a site it cannot measure.

The dimensions

The dimensions — Dimension, Share of web standards, Question
Dimension Share of web standards Question
Discoverability37%Can an agent reach the homepage and find the pages?
Content54%Can an agent read the page without running JavaScript?
Crawler access9%Do crawlers know what they may fetch?
API and agentOnly when the scan finds itIs the API or tool server described for agents?
CommerceOnly when the scan finds itCan an agent read products, prices and policies?

How it works

  1. Fetch public paths only. The scan requests 24 paths on every site, and up to 82 when more checks apply. It also requests a capped set of URLs the site's own documents name. Each response is read to at most 3 MB, and no JavaScript is executed.
  2. Run the 60 checks against what was fetched. 36 are scored and 24 are unscored. No check makes a request of its own.
  3. Score what counts, in two parts: the score is 70% web standards and 30% agent conventions. In each part an essential check counts 3 times its weight, and the part is the share of points earned over the checks that apply. Only a web-standards check holds a level, and an unscored check earns no points.
  4. Fingerprint the result with SHA-256. The fingerprint binds the host, the rubric version, the site profile, and the module scope. It also binds each scored check's id, status, points, counted flag, and exclusion, mcp-server's discoveredBy field, and each unscored check's id and state. /verify recomputes the score and level from the bound outcomes, and reports a mismatch as tamper evidence.

It does not bind the measured origin, the redirect, the defect flag, and every detail, evidence, and fix text. An edit to any of these is not caught.

This site's own surfaces

This site publishes the surfaces below. Each is served now and generated from one content source. It does not pass every check in its own rubric.

/llms.txt llms.txt — Structured summary for language models
/llms-full.txt llms-full.txt — Every page in the page registry and every learn article, in one markdown document
/AGENTS.md AGENTS.md — Build, test and contribution contract for coding agents
/sitemap.md sitemap.md — Every page as a markdown link list
/sitemap.xml sitemap.xml — Machine sitemap with lastmod
/index.json index.json — Typed index of every page and surface
/rubric.json rubric.json — The complete rubric, machine-readable
/robots.txt robots.txt — Named AI crawler rules plus Content-Signal
/ai.txt ai.txt — Plain-language access declaration
/auth.md auth.md — How an agent authenticates (and what is public)
/ai-policy AI policy — Crawling, training, commercial use, contact
/bot Agent policy — Bot identity, rates and contact
/privacy Privacy — What is stored, how long it stays, and what is not collected
/terms Terms of use — What a measurement is, what the API allows, and what the badge licence grants
/abuse Abuse and removal — What the scanner requests, how to block it, and how to request removal
/status Status — The live health of this server: its build, the rubric version and the record store
/.well-known/agent-permissions.json agent-permissions.json — Allow, deny and human-approval rules
/.well-known/mcp/server-card.json MCP server card — The MCP server at /mcp: its name, its version and the protocol versions it speaks. A card names no tool; tools/list does. Its remote is an absolute URL on the canonical origin, so a scan of any other host reports it as declared rather than observed. The MCP server card is an unscored check
/.well-known/api-catalog api-catalog — RFC 9727 catalog of callable APIs. Its targets are absolute URLs on the canonical origin, so a scan of any other host reports them as declared rather than observed, and the scanner scores it partial
/openapi.json openapi.json — OpenAPI description of the API endpoints that answer today
/.well-known/agent-skills/index.json agent-skills index — 59 publishable skill packs covering all 60 checks
/.well-known/security.txt security.txt — RFC 9116 vulnerability disclosure
/.well-known/dns-aid.json dns-aid.json — The DNS-AID records this domain declares, documented in this file and not yet in the zone
/changelog.xml changelog.xml — Atom feed of the rubric changelog, newest first
/.well-known/ard.json ard.json — ARD manifest: one entry, the MCP server card, with the queries it answers
/.well-known/ai-catalog.json ai-catalog.json — AI catalog that lists the MCP server card, for MCP clients

Questions

What is agent readiness?

It is whether an autonomous agent can find your site, read it, trust it, and act on it without a human in the loop. That is narrower than SEO and different from AI visibility: it measures the machine-facing contract, not your ranking or your brand mentions.

How is this different from Lighthouse?

Lighthouse runs in a browser and audits one page at a time, its agentic browsing audits included. This scanner uses a plain HTTP client and no browser.

The scan requests 24 paths on every site, and up to 82 when more checks apply. It also requests a capped set of URLs the site's documents name. It scores what it observed against a published rubric and publishes a record with a fingerprint. A site can use both.

Why does a blocked site get no score instead of a low one?

A site that returns 403 or 429 to every automated client has told us nothing about itself. It has told us something about our vantage point. So the result is no score, with an explanation, rather than a number.

Will it tell me to fix things that do not apply to my site?

No. The scanner decides what kind of site it is looking at before it scores anything. A check whose subject is absent from the site is reported as out of scope. It leaves the score, it is never a failure, and the report says why.

Can I verify a score someone shows me?

Yes, within limits: every record carries a fingerprint. The fingerprint binds the host, the rubric version, the site profile, and the module scope. It also binds each scored check's id, status, points, counted flag, and exclusion, mcp-server's discoveredBy field, and each unscored check's id and state. /verify recomputes the score and level from the bound outcomes, and reports a mismatch as tamper evidence.

It does not bind the measured origin, the redirect, the defect flag, and every detail, evidence, and fix text. An edit to any of these is not caught. Records measured under an earlier rubric are not shown, and a re-scan measures the site under the current one.

Current state

The rubric is published as /rubric.json, the surfaces above are served, and the scanner runs. This site's own score is read from its scan record, as any other site's is.

When a record of this site is stored, it is at /aws/<host> for this site's host, with a markdown copy at /aws/<host>/status.md. No score is published for this site without one.

Scan, fix, re-scan

Every defect lists what the scanner observed and, where one exists, a command that reproduces it. Where a cause can be fixed, the report shows the change to make.

Make the changes, then press Re-scan now on your report. Entering your domain again within 6 hours opens the stored report instead.

To show your score on your own site, press Get the badge on your report and paste its snippet into your homepage. Press Check my badge, and the directory lists your site as a verified listing.

Scan your site