The problem this solves
Before an assistant cites anyone, it resolves entities: what this company is, what category it belongs to, what its product does, how it relates to the alternatives. That resolution leans on high-authority corpora - encyclopedic sources, company databases, structured data - and for many brands those sources are wrong in small, compounding ways. An outdated category label, a stale product description, three conflicting founding stories across three databases.
Entity confusion has a quiet cost: assistants hedge. A brand the graph understands cleanly gets recommended by name with a confident description; a brand with conflicting records gets vaguer treatment or dropped from the shortlist entirely. The damage is easy to miss because each individual record looks almost right, and nobody reads them side by side. Fixing it is unglamorous correction work across sources you do not control directly, which is why it never happens without a map of what is wrong, where, and which route actually corrects each record.
How the mission runs
- Establish the canonical facts. The mission first fixes what should be true - name, category, one-line description, founding facts, product framing - from your own site and input. Corrections need a canon to correct toward.
- Survey the entity sources. The corpora assistants resolve entities from are read for your brand: encyclopedic and database entries, company registries, review-platform categorizations, and your own structured data. Every claim about you is captured with its source.
- Diff against canon. Each captured claim is compared with the canonical facts: correct, stale, conflicting, or missing entirely. Assistants are also asked directly who you are, which shows how the current graph state actually renders in answers.
- Deliver the correction plan. The plan lists every fix with its mechanism: edits you can make directly, structured-data changes for your own site, and correction routes for third-party sources, ordered by how much each source feeds the assistants.
The prompt
This is the exact objective the agent receives. Swap the obvious placeholders for your own domain, segment or channel and run it as-is from the console, Slack, or the API.
What comes back
An entity audit with the full claim inventory - what every consequential source says about you, diffed line by line against canon - plus how the assistants currently describe you verbatim, and a correction plan where every fix names its route, its owner, and its expected lag. Six weeks of unglamorous edits later, the machines describing your company all say the same true thing, and the hedging in answers about you visibly drops.
Make it yours
- Run it before a rebrand or repositioning to inventory everything that will need updating, then again after to verify the graph caught up.
- Include founder and executive entities where personal credibility carries the brand in your category - assistants often resolve a young company through its people before they trust the company record itself.
- Add your top competitor for contrast: seeing their cleaner entity state is often what finally funds the correction work.
Frequently asked questions
Can I even edit these sources?
Directly for some - your structured data, claimed database profiles. Through prescribed processes for others, and encyclopedic sources have community rules the plan respects: it maps the legitimate route for each fix rather than pretending you control everything.
How much does entity state really affect answers?
The audit shows you empirically: assistants are asked about your brand during the survey, and their hedges and errors usually trace to specific conflicting sources. That verbatim evidence, next to the offending records, is the clearest demonstration of the mechanism.
How long until corrections show up in answers?
Structured-data fixes can reflect in weeks through retrieval; encyclopedic and database corrections propagate more slowly and matter most for the next model refresh. The plan flags each fix with its expected lag so nobody re-litigates the work in month two.