How this wiki works

Search engines and AI assistants change quickly, and much of what is written about them repeats unverified claims. This wiki tries to make it clear, sentence by sentence, how we know each thing it says. This page explains the rules.

Evidence levels

Each factual statement is shown in a claim block labelled with one of five levels. An entry’s overall level, shown under its title, is only a summary. The claim blocks are what count.

Level	Meaning	Typical source
Officially documented	The platform says so in its own documentation, policy or announcement.	Google Search Central, Bing Webmaster Guidelines, OpenAI docs
Research-backed	A study with a published method supports it.	Peer-reviewed papers, preprints with data
Observed in practice	Someone measured it, but the platform doesn’t document it. It may differ for your site and may already have changed.	Industry studies, crawl logs, our own re-tests
Hypothesis	A plausible explanation that hasn’t been tested.	Reasoning from documented behaviour
Disputed	Credible sources disagree, or a re-test contradicted it.	Conflicting studies or statements

Officially documented means the platform said it, not that it is true in every case. Where documentation and observation disagree, the entry shows both.

Claim checks

Popular claims get their own entry (claim check) with a verdict: supported, partly supported, not supported, false, or unverified. Unverified means there isn’t enough evidence either way. It is not a polite way of saying false.

Sources

• Every claim names its source, and each entry lists its sources at the end. Sources are kept in a single registry, so the same document is cited the same way everywhere.
• Links are checked weekly. A broken link, a redirect to another page, or a page whose title no longer matches is reported.
• Source changes are detected. We fingerprint each source’s main text. When it changes, the entries citing it are listed for re-review.
• Quotations are verified. A quoted passage must appear verbatim in the source it cites. A weekly check fetches the sources and flags any quote it can’t find.
• Unpublished reports (industry studies shared as PDFs, not publicly available) can be cited, marked as such, with no link. Readers can’t check them, so a claim resting only on them can be no stronger than observed. Change logs never cite them.

Dates and freshness

• Each entry shows when a person last reviewed its facts against sources. This date is the last check, which may differ from the last edit.
• Each kind of entry has a review interval (from 90 days for change logs to a year for concepts). Past it, the page says it may be out of date.
• Observations carry the month they were seen. After nine months, the claim shows a notice asking readers to re-test before relying on it.

Re-testing observations

Some observations are tied to a fixed set of prompts that can be run again. Examples include how often ChatGPT cites third-party sites, and how many runs are needed to measure a brand’s visibility. A re-test keeps every raw answer and source, and its result is shown under the claim.

• Consistent: the observation date is renewed.
• Differs: the claim is flagged until the entry is rewritten or the claim is marked disputed.
• Inconclusive: recorded, nothing changes.

Prompts in a set are never edited after the first run. A changed set gets a new name, so results stay comparable. Assistants that can’t be queried automatically, such as Chinese consumer apps, are tested by hand.

Translations

English is the reference version. A Chinese translation must contain the same claim blocks, in the same order, with the same levels and sources. It records which English review it was translated from. When the English entry is reviewed again, the translation says it may be behind and links to the English text. Pages not yet translated are shown in English and labelled.

What we can’t verify

• How any platform ranks or chooses sources internally. We report what platforms document and what can be observed from outside.
• Observations are samples: from one time, place, account and set of prompts. They show that something happens, not how often it happens everywhere.
• Unpublished reports, which readers can’t check.

How entries are written

Entries may be drafted with AI assistance. Every statement must still pass the same checks: an evidence level, a registered source, and verified quotations. Machine-readable versions of every entry are published as Markdown and listed in /llms.txt.