What is a trust graph?

The idea

A model does not hold a ranked list of companies. What it holds is a dense web of associations built from everything it has read: which brands are discussed alongside which problems, who is described as credible by whom, which claims are corroborated and which appear only in marketing copy.

When it answers a question, it draws on that web. Trust graph is the useful name for it — useful because it makes the right thing obvious: your position is determined by relationships between sources, not by any property of your own site in isolation.

What feeds it

• Training data. Everything published before the cutoff, weighted by how widely and consistently it was repeated.
• Retrieval sources. What the assistant fetches at query time, which is where recent and specific information enters.
• Third-party evaluation. Reviews, comparisons, analyst coverage, community discussion — the material that reads as independent.
• Brand-supplied structured data, where the surface accepts it.
• Corroboration across all of the above. A claim repeated consistently by unrelated sources behaves very differently from the same claim made once.

Why owned content has limited leverage

This is the part most brands find uncomfortable. Your own site is one source among thousands, and it is the one the model has the most reason to discount, because every brand describes itself favourably.

The signals that move a trust graph are largely ones you do not control directly: whether independent publications cover you, whether users discuss you in communities, whether comparison sites list you, whether what all of them say is consistent. That is closer to public relations and product reality than to content marketing.

What your own content can do is be accurate, specific and consistent enough to be safely repeated. A model that finds precise, checkable claims on your site and the same claims corroborated elsewhere will use them. One that finds vague positioning language has nothing to lift.

The failure mode to avoid

Inconsistency is worse than absence. If your pricing page, your sales collateral and a third-party review disagree about what you cost, a model reconciling them will hedge, caveat or omit you entirely rather than pick one.

This is also the mechanism behind hallucination liability: where the available information is contradictory, the model resolves it by generating something plausible, and what it generates may be wrong in ways that damage you. Consistency across every source that mentions you is the cheapest and most reliable investment available.