Why do entity inconsistencies cost citations?
Because grounding is a trust calculation performed at speed. When an engine considers quoting your facts, it cross-checks who you are across the sources it can see. If your site says “founded 2014”, a directory says 2016 and your LinkedIn page says nothing, the engine does not average them — it downweights the whole entity. Contradiction reads as unreliability, and unreliable entities get fewer citations for everything, including their true claims. The mechanism is detailed in entities and the knowledge graph and grounding fact consistency.
Minutes 0-5: Define the canonical record
Open a blank note and write, exactly, the canonical facts: legal name and trading name; founding date; headquarters address; one-sentence description (the version you want engines to repeat); official domains and social handles. This note is the source of truth everything else is audited against. If the room argues about the one-sentence description, good — that argument was the audit’s first finding.
Minutes 5-12: Check your own site
- About page, footer, contact page: do name, date, address and description match the canonical record byte-for-byte?
- Organization schema: name, alternateName, foundingDate, address and a complete sameAs set pointing at every official profile. View-source the rendered HTML — schema injected client-side may never reach crawlers. See structured data for AI.
- Press or news pages: old taglines and superseded descriptions linger here for years.
Minutes 12-20: Check the profile layer
Open each profile the engines can see — LinkedIn, Crunchbase, review platforms, key directories, Google Business Profile if local — and compare against the canonical record. Common cracks: a founding date that differs by a year, an old address on a directory nobody owns, three descriptions from three eras of positioning, a former brand name still live on a forgotten profile. Log every mismatch with a URL; you will fix them in order of citation importance, not alphabetical.
Minutes 20-27: Ask the engines
The direct test: ask ChatGPT, Google’s AI surfaces and Perplexity the entity questions — “who is X”, “when was X founded”, “what does X do”, “where is X based”. Record every answer. Watch for three failure types: wrong facts (trace them to the source you logged above), entity confusion (you are merged with a same-named company — the hardest problem, requiring disambiguation work), and vagueness (no confident answer at all, meaning the consensus is too thin). Screenshot everything into your evidence archive.
Minutes 27-30: Write the fix list
- Site fixes: align about page, footer and schema to the canonical record — usually same-day work.
- Profile fixes: claim and correct the mismatched profiles, starting with the ones engines cited in your test answers.
- Escalations: entity confusion or a missing Wikidata item goes to the entity workstream, with the honest-anchors approach from Wikipedia notability in 2026.
- Process: add “update every profile” to the checklist that runs whenever brand facts change — new address, new tagline, new round.
How often should you run it?
Thirty minutes quarterly, inside the quarterly GEO audit, plus ad-hoc after any brand fact changes. Multilingual brands run it per language — cross-language entity fragmentation is its own failure class; see the multilingual rollout walkthrough. The full 50-point context is in the GEO checklist. Keep the canonical record in shared documentation with a named owner; a canonical note without an owner drifts within two quarters, and the audit quietly stops meaning anything.
The audit works because it matches how the engines work: they do not know who you are, they compute it. Thirty minutes of making the computation easy is the cheapest citation defense in the discipline.