AI-Readiness Score · Anchor Data Labs
Can AI actually read
your help center?
AI assistants answer customer questions from whatever they can read of your docs. This scan reads your help center the way an AI does, generates plausible customer questions from your content, and shows you with evidence where those answers aren't supported by what you've published. Free.
Scanning your help center…
- Finding your articles
- Reading it the way an AI does
- Asking plausible customer questions
- Writing up the evidence
Usually 2–5 minutes. You can close this tab; we’ll email your report link, and tell you if the site can’t be scored.
- Evidence, not vibes. Every finding shows the page the AI retrieved and the passage it surfaced, and tells you when it retrieved nothing.
- Scored like an AI reads. Seven dimensions, 0–100, tested with the kinds of questions your customers ask.
- Yours to share. A permanent report link + a share card for your team.
What this scan actually does
We run your published help center through a reference retrieval-and-answering pipeline, the same shape as the systems that answer questions about you. It crawls your public pages, splits them into passages the way a retrieval system does, generates plausible customer questions from your content and business context, and checks whether an AI answering only from your content gets them right. Every finding names the page it came from and shows the passage.
The seven dimensions
Each is scored 0–100; the site score is their weighted mean.
- Retrieval performance25%
Can an AI find and give the right answer to a real customer question from your content?
- Structure & chunkability15%
Is each page split into clean, self-describing sections a retriever can isolate?
- Self-containment15%
Does a passage stand on its own, or lean on "as mentioned above" and undefined acronyms?
- Contradictions15%
Do two pages give conflicting answers to the same question?
- Freshness10%
Is the content current, or dated in ways that make an AI distrust it?
- Topic coverage10%
Are the topics your customers ask about actually present?
- Machine accessibility10%
Can AI crawlers reach the pages at all, per your robots.txt and how the pages render?
Scores roll up to a 0–100 number and a grade: AI-READY at 85 and up, SERVICEABLE at 70, AT RISK at 50, FAILING below that.
A sample finding
A Billing FAQ that says refunds take 14 days while the Refund policy says 30 days is a contradiction. The report names both articles, quotes both sentences side by side, and marks it high severity, because an AI answering a refund question may give inconsistent answers or cite the wrong policy. You can verify every finding yourself in seconds.
What it reads
Public HTML help centers and documentation sites. It reads a page the way a crawler does, and falls back to a rendered read for JavaScript-heavy pages. The scoring is tuned for English content today. Pages in other languages are detected and set aside, and a site without enough English pages is reported as degraded rather than mis-scored.
Sampling and limits
It samples rather than exhausts: up to 200 pages, around 50 customer questions, and up to 20 article pairs checked for contradictions. Because it uses a language model to read and answer, results can vary slightly between runs. If the model or its provider is briefly unavailable, those questions are marked "could not assess" and left out of the score, never counted as failures.
What the score does not prove
It is a diagnostic signal rather than a certification. A high score does not guarantee an AI will answer every question perfectly, and a low score is not a verdict on your product, only on how legible your published content is to an AI reading it. It does not use your real customer-question logs, and it does not measure any specific vendor's assistant.
Want a human read? The scan tells you what is off. The human-reviewed diagnostic is a person going through your findings by hand and handing you a prioritized, do-this-first roadmap.