Methodology · v0.3
How the audit works, in full
Every check, what it looks for, and how much weight its result deserves. Published because a score you cannot interrogate is not worth much.
What gets audited
We fetch your homepage and up to four automatically selected representative pages — typically an About, contact, product or service, and article or resource page. We read the HTML your server returns, without executing JavaScript, because not every automated visitor does.
Categories and weights
| Category | Weight |
|---|---|
| Technical AI access | 30% |
| Content readiness | 25% |
| Entity confidence | 20% |
| Content trust | 20% |
| Freshness | 5% |
A category with no applicable evidence is marked unavailable and excluded from the calculation rather than scored as zero, so a missing signal does not silently penalise a site.
Observed and heuristic checks
Checks are labelled by how directly the result was established, and the distinction matters when you read a report.
- Observed (14 checks) — read straight from the response. A status code, a robots directive, the presence of a tag. If it is reported, it is what your server returned.
- Heuristic (6 checks) — inferred by pattern matching against page content. A heuristic failure means “not found the usual way”, which is a weaker claim than “absent”. Each states how it can be wrong.
Every check
20 checks, grouped by the confidence they carry. The identifier is stable across versions so you can track a specific check over time.
| Check | Confidence | Why it matters |
|---|---|---|
page-response | observed | If the homepage does not return a successful response to an ordinary request, nothing else about the site can be assessed by an automated visitor, and neither crawlers nor AI systems can read it. |
https | observed | HTTPS is a baseline trust and integrity signal. Pages served over plain HTTP can be modified in transit, and many tools treat them as lower quality or refuse to follow them. |
robots | observed | robots.txt is where a site states which automated visitors may access what. When it is missing or unreachable, crawler behaviour is left to each crawler's own defaults rather than your stated intent. |
search-crawler-access | observed | Search and citation crawlers fetch pages to answer live questions and cite sources. Blocking one at the site root means content cannot be retrieved for those answers. Allowing them removes that specific barrier — it does not make the site appear in any answer, and no directive can guarantee that. How this can be wrong: Reads the stated policy only. It does not test whether any crawler actually visited the site, nor whether a crawler honours what it was asked to do, and no robots directive determines whether a site is cited. |
training-crawler-policy | observed | Crawlers that collect data for model training are a separate matter from those that retrieve pages to answer questions, and many publishers restrict one while permitting the other. That is a licensing decision about how the work may be used, so this check reports what the policy says and deliberately does not score it — a site that blocks training access has not made a mistake. How this can be wrong: Reads the stated policy only. Some documented training crawlers have a history of ignoring robots.txt, so a stated restriction is a request rather than an enforcement. |
sitemap | observed | A sitemap lists the URLs you consider worth reading, so a crawler does not have to discover them by following links. It matters most for large sites and for pages that are not well linked from the homepage. |
representative-coverage | observed | Reports how much of the sampled page set was actually readable. Low coverage means this report rests on fewer pages, so treat its site-wide conclusions as weaker rather than as a finding about the site itself. |
organization-schema | observed | Organization structured data states, in a form machines read without guessing, who the site belongs to — the name, and identifiers linking it to other profiles. Without it, a machine has to infer your identity from prose, and inference is where the wrong company, or a competitor, gets named instead. |
about-link | heuristic | An About page is the usual place both people and machines look to confirm what an organization is and does. Its absence removes the most conventional evidence of legitimacy a site can offer. How this can be wrong: Detected by matching link text and URLs on the homepage. About information reached only through a menu, a footer image, or an unusual label may exist but not be found here. |
contact-link | heuristic | Reachable contact information is a long-standing accountability signal: it shows a real party stands behind the content. Its absence is common on low-quality and scraped sites, which is exactly the company you do not want to keep. How this can be wrong: Detected by matching link text and URLs on the homepage. Contact details shown only as an image, inside a form widget, or on a page linked unusually may be missed. |
site-title | observed | The page title is the shortest statement of what a site is, and it is used heavily when systems summarise or list a page. A missing, generic, or template-default title wastes the most prominent naming opportunity a page has. |
meta-description | observed | The description is a supplied summary of the page. When it is absent, systems that need a summary generate one from whatever text they find, which is rarely the framing you would have chosen. |
h1 | observed | One clear primary heading states what a page is about and anchors the structure beneath it. Missing or competing top-level headings make it harder to determine which topic the page actually addresses. |
canonical | observed | A canonical URL names the one address that should represent a page when the same content is reachable several ways. Without it, attention and references can be split across duplicates instead of accumulating on one URL. |
substantive-copy | observed | A page needs enough readable text to answer something. Pages that are mostly navigation, imagery, or text assembled by scripts after load give a reader very little to work with, however good they look to a person. How this can be wrong: Measures text present in the served HTML. A page that renders its content with JavaScript may read as thin here while appearing complete in a browser — which is itself worth knowing, since not every automated visitor executes scripts. |
unique-titles | observed | Distinct titles let each page be told apart. Repeated or templated titles across pages make a site look like near-duplicate content and obscure what any single page uniquely offers. |
authorship | heuristic | Naming who produced or is responsible for content is one of the clearest accountability signals available, and one of the hardest for low-effort content to fake convincingly. Its absence is a frequent marker of unattributed bulk content. How this can be wrong: Detected by looking for author markup and common bylines. Authorship credited only in an image, a bio block, or an unusual layout may be present but not detected. |
external-sources | heuristic | Linking to sources shows claims rest on something checkable. Content that cites nothing is harder to corroborate, and corroboration is much of what determines whether a claim gets repeated. How this can be wrong: Counts outbound links to other domains. It does not judge whether those sources are relevant or credible, and a page citing named sources without linking them will read as unsourced here. |
privacy | heuristic | Privacy and policy pages indicate an operator accountable for how the site handles people. Their absence is unusual for an established organization and conspicuous on sites built to be disposable. How this can be wrong: Detected by matching link text and URLs. A policy reached only through a cookie banner, a script-rendered footer, or an unusual label may be missed. |
freshness | heuristic | A visible date tells a reader whether information is current. For anything time-sensitive, undated content cannot be judged as still true, so it tends to be treated cautiously — or passed over for something dated. How this can be wrong: Looks for date markup and visible date text. A page kept current without ever showing a date will read as stale here, and a displayed date is a claim by the site, not proof of a recent update. |
The score is rule-based
No language model produces any score, finding, or verdict. Checks are deterministic rules, so the same site produces the same result on consecutive runs and any finding can be traced to the rule that generated it. Explanatory text — including everything on this page — is written by hand and keyed to a check identifier.
What this does not tell you
- The score is ours. It is not used by any AI platform and is not an official ranking factor anywhere.
- It measures readiness, not observed visibility. We do not currently query AI systems to see whether they mention your brand, and we do not report that we do.
- Improving it removes barriers to being read and understood. It does not guarantee citation, and nothing does.
- Accuracy against a human-labelled benchmark is still being established. Published completion and coverage figures describe whether audits ran, not whether every finding was correct.
Versioning
Every report records the methodology version that produced it. Changes to checks or weights bump the version, because scores from different versions are not directly comparable.
Read it against your own site
Every finding in a report links back to the check that produced it.
Run a free audit