How HonestClaw scans and scores agent readiness

Use means published signals. We do not claim an agent finished the product task. Live API or MCP call success, and tool quality when called, stay out of the scored number today.

The score measures whether useful information and capabilities are published clearly enough for agents to find and read. It does not score whether a capability is cheap, fast, or generous.

A Free Scan produces the public Agent Readiness Score.

Free Scan reads the version of a page the server sends directly, without running the site's JavaScript. That is one defined retrieval model, not a claim about every AI agent.

The scan

• never logs in
• never breaks access controls
• never runs a live product task through to completion

Each score is a company-level reading on the main company domain (the registrable domain, such as stripe.com). Subdomains and pasted URLs are treated as part of the same company.

In scope are public surfaces the company publishes or points to as part of its product surface, where the scan can retrieve them without logging in:

• pages
• robots.txt
• sitemaps
• well-known endpoints
• linked docs
• official developer or documentation surfaces
• other discovery files as published

Capabilities that exist only behind login sit outside this measurement unless they are published where an outside agent can retrieve them.

The public score is calculated only from checks where we reached a real answer (conclusive scored checks). If we cannot tell, we leave that check out of the score calculation (it exits the denominator). It lowers confidence. It does not invent a pass or fail.

Confidence is evidence coverage: how many scored checks reached a real answer, divided by total scored checks.

Example: If 40 of 45 scored checks reached a real answer, evidence coverage is 89% and confidence is High. If only 20 reached an answer (44%), we don't show a headline score, because most of the site could not be assessed.

We distinguish "we checked and it wasn't there" from "we couldn't tell."

The Agent Readiness Score uses the same set of capabilities for every site. It is not adjusted by company category. We call this a universal capability index.

A site without APIs or docs may score lower even when those surfaces are not central to its business.

• Citation visibility in answer engines stays out of the Agent Readiness Score.
• Payment-protocol checklist items stay out of the scored number.
• There are no bonus points outside the scored checks.

A deeper look is private and account-owned. The product name for that paid path is Deep Scan. Free and paid use the same scoring engine. Payment buys depth and workflow.

Deep Scan privately looks across more of the site and can collect evidence that Free Scan could not reach, including JavaScript-rendered content. It never rewrites the public score, confidence, or rank.

We do not sell scores or placement.

Every site scanned under one scanner version gets the same checks, the same fixed scoring rules, the same scanner profile, and evidence for each check. No evaluator model writes the score, and Agent Comprehension is display-only. ARS gives no credit for workarounds: it scores what the site exposes, not what a capable agent might manage without it.

The same site state, scanned with the same version, is designed to give the same or nearly the same result. Each public result shows its scanner version.

Task-based scores tell you how one agent setup did on tasks someone chose. They do not compare sites like for like, because:

1. The agent is part of the result. The model, its version and its tools change over time, so a site's result can move when nothing on the site changed.
2. Task choice shapes the score. A different task set or reworded prompt can give a different result on the same site.
3. Runs vary. The same agent can pass a task once and fail it the next time (τ-bench, Yao et al., 2024). A steady pass rate needs repeated runs.
4. Logins, CAPTCHAs and payments block testing. Getting past them on many sites is hard to do safely or consistently, and live data changes between runs.
5. A pass rate does not say what to fix. ARS reports each check with what the scanner found.

Use task testing to see whether a given agent completes given tasks. A site can do well on ARS and poorly on a task test, or the reverse.

All 42 scored checks look at what a company publishes for its own domain: its site, its subdomains, its DNS records, a status page it links to, and spec files it points to on other hosts. ARS does not score appearances or citations in AI answers (the target of AEO and GEO), brand awareness, what models already know about a brand, search rankings, or directory listings.

Those outcomes depend on things outside the site: the model and its training, the question, competing pages and the date. If they counted, two sites exposing the same things could score differently only because a model knows one brand better. A higher ARS is not a claim that a site will be shown or cited.

Check the profile and its version, which checks or tasks ran, and whether a model wrote the number. If a score includes visibility or citations, check which questions, answer engines and dates it used.

• Every scored finding carries the evidence the scanner observed.
• Checks the scanner could not resolve drop out of the score and lower its confidence. Confirmed absence can lower the score.
• Confidence is the share of scored checks that resolved.
• Method changes bump the scanner version. Scores are not hand-edited.
• Published research supports the direction of each check, not HonestClaw's exact point weights.

These are separate claims. ARS is designed to be stable (same site state and version, same result) and comparable within one version. It is designed to measure only the site, by keeping models, popularity and visibility out of the score. We have not yet published test-retest numbers or a study of whether scores track agent outcomes.

Under scanner 1.69.0, a site at the top level of every scored check scores 100. Scores from scanner versions before 1.68.0 used a different scale. On those versions, a site with every check resolved could score at most 93 (1.13 to 1.67) or 94 (1.11 to 1.12), so compare a score only with scores from the same scanner version.

The score is an observation of published surfaces at a point in time.

It is not a

• security audit
• credit rating
• product review
• endorsement
• measure of whether agents prefer, cite, or buy from a company

It also does not measure live task completion.

When Agent Comprehension appears on a result, it shows clarity of readable facts from pages the scan reached. It is not folded into the Agent Readiness Score. More on Features.

• Weights are product judgment, not yet calibrated against agent outcomes. About 55% of points are in developer API, authentication and API documentation checks, so sites without a public API score lower even when they don't need one.
• The leaderboard can currently lag. Use the score on the site's own result page.
• Read each score with its confidence level. A site with more unresolved checks can score higher than it would if those checks had been found missing.
• The scanner is one identified HonestClaw client whose user-agent includes ChatGPT-User, Claude-User and Perplexity-User. Pages that vary by sign-in, location or account are scored as served to it, so a score is not a claim about what every agent sees.
• Some checks overlap (one OpenAPI file can lift several), one check penalizes page metadata older than 730 days, and some early draft standards are scored before they settle.

Each check asks one specific question about the site.

The better the evidence we find, the more of that check's available points the site earns. If we cannot reach a reliable answer, we mark the check Unknown and leave it out of the score rather than treating it as a failure.

For each check where we reached a real answer, points earned equal that share times the check's nominal maxPoints. A check that cannot reach High confidence still earns against that nominal max. The denominator uses scoringMaxPoints, the most the check can actually contribute. The check reference lists that earnable cap as the effective max.

Score. Agent Readiness Score = 100 × points earned ÷ scoringMaxPoints across conclusive scored checks, rounded to the nearest whole number.

Scored findings include the evidence behind them. The public score is computed from versioned check results and scoring rules. Model-assisted analysis does not overwrite the public check states or score.

Informational checks can appear on the result without entering the readiness number. Evidence density and related measurements stay informational until they are calibrated on their own.

Check maxPoints, scoringMaxPoints, which checks are scored, and the plain criteria for each state live in the named versioned check reference for that scanner version. With that page and the site's public scan results, anyone can redo the math behind a published score. The number of checks can change between scanner versions.

Weights and max points are versioned product judgment. Each check lists the sources behind it, where it has any. Weights have not yet been calibrated against measured agent outcomes, and they are not scientific constants.

The public result groups checks into dimensions.

Robots baseline. When robots.txt blocks a needed surface for a generic compliant agent (the * baseline), that check resolves as conclusive Not detected (not_detected). The surface is treated as unavailable to that agent class, and the zero stays in the score.

Could not verify. When temporary server errors, timeouts, network failures, access gates that prevent assessment, or scan budget exhaustion leave the scanner without enough evidence for an honest conclusion, the check resolves as Unknown (unknown). Responses that conclusively establish presence or absence can still produce a conclusive state. Unknown is left out of the score calculation and lowers confidence. Incomplete evidence is not treated as absence.

On Free Scan, content that only appears after JavaScript often records as Not detected, because the free path never runs site JavaScript. That is different from Unknown, which means the scan could not reach a reliable answer.

Agent standards are still evolving. We distinguish established signals from newer conventions and record that distinction in the versioned check reference. Established protocol and support signals can be scored; emerging declarations and conventions may appear as informational or differently weighted checks. llms.txt is a newer convention that is scored, at up to 5 of 160 points, because it is a file an agent can read directly; the score does not claim it affects AI-answer citations.

Every score is a point-in-time reading. It is bound to the scanner version shown on the public result.

Score-affecting changes bump the scanner version. We keep the scoring rules for prior scanner versions. Re-scan after the site changes. Method or coverage changes can change a later reading.

A correction comes from a new scan, or from a versioned methodology or software fix that ships under a new scanner version.

Write to disputes@honestclaw.ai. We correct errors. We leave an accurate result in place when it is unflattering. Opening a dispute is not a purchase.

• For a practice checklist, read Make your site readable by agents.

Run a free scan from the homepage.