How Docsbook proves what it claims
Documentation platforms are sold with adjectives — advanced, intelligent, optimised. Adjectives are unfalsifiable, so they carry no information. These docs are written to a rule instead: a capability page must let a sceptical engineer check us. If a sentence cannot survive the question "says who?", it does not ship.
This page is the rule itself, so you can hold us to it.
What every capability page contains#
Each page under SEO, GEO, AEO, Agent-ready content, AI chat, Analytics and Translations carries four blocks, in this order.
| Block | What it must contain | How to check it |
|---|---|---|
| What you get | The outcome in your terms — what appears on the page, in the panel, or in the API | Open your own site and look |
| How it is built | The mechanism at the level only someone who read the implementation could write: thresholds, order of preference, fallbacks, what happens on failure | Compare against the rendered HTML, the API response, or the exported data |
| Evidence | Each rule as rule → why the machine or the reader behaves that way → a link to a primary source | Open the link |
| Limits | What the feature does not do, what is not measured, what depends on a model version we do not control | Judge whether we left anything out |
A page with no limits block is a brochure, and brochures are believed less, not more.
What counts as a source here#
We rank sources, and the ranking is visible in how a claim is worded.
- Specifications and vendor documentation — Google Search Central, schema.org, W3C, the IETF, the Model Context Protocol specification, the llms.txt specification, and the published documentation of the model providers involved. A claim resting on one of these is stated flatly.
- Peer-reviewed or preprint research with a stated method and sample size. Stated with the method attached: "measured on 10,000 queries" tells you how far to trust it.
- Independent studies with published methodology. Always attributed in the sentence — "X measured", never "it is known that".
- Vendor-reported results, including our own. Labelled as vendor-reported. Our own internal measurements are labelled as ours.
Anything below that line is not evidence, and does not appear.
What we deliberately do not claim#
These are the claims a documentation vendor is expected to make, and the reason we do not make them.
- No citation-rate lift. Nobody can honestly promise that turning on a feature makes Perplexity or ChatGPT cite you N% more often. What has been measured in controlled work is narrower than the marketing version of it, and GEO says exactly what was measured and by whom.
- No AI answer-accuracy percentage. A number produced on our own corpus, with our own grader, is not evidence about your corpus. AI chat describes what the pipeline does to stay grounded and what it measures, without inventing a score.
- No ranking guarantees. Search ranking is not a contract, and Google says so in its own documentation. SEO separates what is documented by Google from what is inference.
- No compliance status we have not earned. MCP security states the current position and names what is not offered yet.
"Under question" blocks#
When a claim in these docs cannot be verified — the mechanism changed, the source turned out not to say what it was cited for, or nobody has published a measurement — we do not delete the sentence and we do not soften it into vagueness. It becomes an explicit block:
Under question. The claim. What is verifiable: the part that is. What is not: the part that is not, and why. Treat it as a hypothesis until measured.
That block is a promise about our process: a claim we cannot support is visible to you rather than quietly removed.
How these claims stay true#
Documentation drifts away from a product silently, which is the failure mode Docsbook exists to fix — so the same machinery runs on these docs.
- Every user-facing change ships with an entry in the changelog, stating what the change was meant to buy, not just what moved.
- The changelog is projected into per-outcome pages, so you can read the history of one outcome — AI citations, support load, organic traffic — rather than a flat list.
- Pages carry a visible last-modified date taken from the commit that changed them, so a stale page cannot pretend to be current. See GEO.
Found something wrong?#
A claim on these pages that does not match what the product does is a defect, and we would rather hear it from you than have it sit there.
- Email support@docsbook.io with the page and the sentence.
- Or say it in the Docsbook Discord.