Agent Readiness Benchmark: the method

Published 2026-08-26 by ApexClaw Research. Engine ARB/1.1. 49 checks, 9 weighted pillars, every finding carrying the exact string observed on the target so any result can be reproduced or disputed by inspection.

ARB/1.1 measures whether this one public page is legible and independently verifiable to an autonomous agent: structured data, machine-readable dates, off-site corroboration, and the page's public transport and security surface. It does not measure the organisation's actual security posture, the quality of its product, or its reputation -- those require access, judgement, and time this engine does not have and will not simulate. Its security_surface pillar reads public response headers and TLS parameters observed on one fetch; it is NOT a penetration test and must never be read as one.

Rubric authored and reviewed by ApexClaw Research. Engine ARB/1.1, published 2026-08-26, revised whenever a check or weight changes -- full history in Versioning below.

What is agent readiness?

Agent readiness is the degree to which a website can be fetched, parsed, cited and transacted with by an autonomous agent rather than a human reader.

Why measure it separately from SEO?

Search ranking assumes a human clicks a result. An agent never clicks: it retrieves, extracts an answer, and either cites the source or omits it. Omission is a retrieval failure, not a ranking one.

How is a score produced?

Each check returns one of three verdicts: VERIFIED, UNVERIFIED, or UNOBSERVABLE. There is no partial credit -- a check either verifies the fact it checks for, or it does not. An UNOBSERVABLE check (the evidence needed to decide was absent, e.g. a fetch failed or a third-party API was unreachable) is removed from the denominator rather than counted as a failure, so nothing is inferred from missing information.

What does the percentile mean?

The percentile compares a score against 30 real sites measured with this same engine. Median 59.8.

Which checks are measurements and which are inferences?

A deterministic check reads a structural fact: a header, a status code, a JSON-LD type, an element count. A heuristic check infers from a text pattern and can miss a true signal.

What are the limits of a single run?

A run scores one URL and the site-wide files at its root: robots.txt, llms.txt, sitemap.xml and the well-known paths. A finding of absent means absent from that page, not from the site.

How is a disputed finding settled?

Every finding carries its evidence string. A dispute is resolved by fetching the same URL and comparing, not by argument. Re-running after a fix shows the delta attributable to that single change.

Versioning

The rubric is versioned. Weights and checks may change; a published score names the engine version that produced it, and an earlier version's published score is never restated under a later version's rubric. Changes are dated in the machine version.

Why publish the rubric instead of keeping it proprietary?

A benchmark that cannot be reproduced is an opinion. Publishing the checks, the weights and the cohort is what makes a result arguable, and a result that can be argued with is one that can be cited.

The alternative, a hidden score, forces a reader to trust the scorer. That works for an established authority and fails for a new one. So the whole method is on this page, the machine-readable version is one request away, and any figure quoted from either carries the engine version that produced it. A competitor can implement the same rubric and get the same numbers; that is the point rather than a risk. Standards earn their position by being adopted, and nothing is adopted that cannot first be inspected.

What this benchmark does not measure

ARB/1.1 measures whether this one public page is legible and independently verifiable to an autonomous agent: structured data, machine-readable dates, off-site corroboration, and the page's public transport and security surface. It does not measure the organisation's actual security posture, the quality of its product, or its reputation -- those require access, judgement, and time this engine does not have and will not simulate. Its security_surface pillar reads public response headers and TLS parameters observed on one fetch; it is NOT a penetration test and must never be read as one.

Four limits follow from that, and each is stated on the face of every report rather than buried here. A run covers one URL plus the files at the domain root, so an absent finding means absent from that page. 5 checks infer from a text pattern instead of reading a structural fact, and those are tagged as heuristic wherever they appear. A check that cannot be evaluated at all is UNOBSERVABLE and is removed from the denominator instead of being scored as a failure, because counting an unknown as a failure would penalise a site for the benchmark's own blind spot. The security_surface pillar reads response headers and TLS parameters observed on one fetch -- it is NOT a penetration test and must never be read as one.

Sources

Run it

Score any public URL at the audit form. The free result names the score, the pillar breakdown and nine findings; the full report adds the remaining checks, the ranked remediation queue and a JSON export.

Contact hello@apexclawai.com.

---

About ApexClaw · Contact · Privacy · Terms · Agent access policy