The knowledge base

Vercatus does not apply a checklist. Its plays and the site audit’s findings rest on hypotheses from a knowledge base, retrieved where they apply and reasoned with. Each entry records:

• the claim, in one paragraph;
• a grade. A: official documentation or standard · B: peer-reviewed or large controlled study · C: reputable industry data or observational · D: practitioner consensus;
• its sources, with year;
• an effect estimate, direction and size, with the basis it rests on: measured (a study reports a figure), reported (an industry report or announcement gives one), documented (an official source says what happens but gives no size) or unquantified (a direction with no figure). A grade says who is speaking; the basis says how much the effect is known. A documented behaviour is never turned into a forecast;
• applies when: the site conditions under which it holds (informational, e-commerce, multilingual, CDN behind Cloudflare …);
• the engines it concerns;
• a revalidate date, because this field changes quarterly;
• a status: verified (sources located and figures checked) or reported (supplied second-hand and not yet located). Agents are instructed never to quote a reported figure as fact.

The whole base is published at vercatus.com/knowledge, one page per entry, claim first, with its grade, sources and re-check date; plays and findings link to the entries they rest on.

The base holds 192 entries across search-engine documentation, IETF and W3C standards (RFC 9421 Web Bot Auth, RFC 9309 robots, the AI-preferences work, TDMRep), retrieval research (hybrid retrieval, lost-in- the-middle, chunking, reranking), generative-engine research (GEO, AI Overview audits), agentic commerce protocols and what an agent needs to buy or book on a site, spam policies, and dated 2025–2026 industry measurements of AI-search behaviour. Each entry carries a stable rule number (R-001 onwards) that plays and findings cite.

Every finding names the entries it rests on, and the grades appear beside them in the run report. You can browse the whole base in the workspace under Knowledge base.

The observed line

A grade says how strong the published evidence is. It does not say whether the rule works for the brands Vercatus measures. So each rule that a kind of play rests on also carries an observed line, from the verdicts of plays of that kind across sites:

• Supported by our verdicts: plays resting on the rule won more often than not, with their typical effect.
• Not observed: not enough verdicts yet, or fewer than three sites behind them.
• Contradicted: the verdicts point the other way. A contradicted rule is demoted from play generation.

Each observed line says how many verdicts and sites it rests on. It holds aggregates only. See Vercatus Intelligence.