What Makes a Source Trustworthy to a Large Language Model

A language model does not keep a trust score for each source, at least not one any AI company has published. What looks like trust from the outside, as a synthesis of the evidence below rather than any company’s description, combines several observable things: how much weight a source carried in training, how well it ranks when a system searches, how widely other independent sources corroborate it, and how clearly its own content supports its claims. A source that does well on all four tends to appear often and be described consistently. One that does well on none can still be repeated, sometimes wrongly, when someone asks about it by name.

What the platform makers say

The companies’ public documents say little about trust inside their models, but Google says a good deal about trust in the ranking systems that feed its AI features. Its guidance on creating helpful content says its systems look for signs of experience, expertise, authoritativeness and trustworthiness, and that “of these aspects, trust is most important.” Its AI optimization guide says AI responses retrieve pages through Google’s “core Search ranking systems.” Google anchors AI Overviews to its results: Liz Reid of Google wrote, in a May 2024 post on AI Overview errors, that they “are built to only show information that is backed up by top web results.”

For Google’s AI features, then, a trustworthy source is largely one that Google’s ranking systems already treat as trustworthy.

For training, a clear public statement is older. OpenAI’s GPT-3 paper explains that datasets “we view as higher-quality are sampled more frequently.” Sources judged higher quality left a larger mark on that model. That is a 2020 description; it shows the principle, not today’s exact mix.

Corroboration: the difference between real and fake experts

Useful evidence about what makes a source credible to AI models comes from an experiment with fake ones. Authoritas tested 11 fake expert personas, exposed after being featured across more than 600 UK media articles, on nine AI models. Asked 55 topic questions about leading experts, the models never named the fake personas. The study contrasts them with real authorities who have built “deep, corroborated brand entities,” with consistent profiles and third-party validation.

The lesson is that many mentions were not enough. What the fake personas lacked was independent corroboration: other credible sources confirming who they were and what they knew. Google’s guide makes a related point for its own systems: inauthentic “mentions” do not help.

Why some brands appear consistently

AI answers vary a great deal, yet some brands keep turning up. In a January 2026 SparkToro study of AI brand recommendations, the lists themselves almost never repeated. Within that variation, the leading brands in a category still appeared in a large share of answers: across 994 responses about headphones, brands such as Bose, Sony, Sennheiser and Apple showed up 55% to 77% of the time.

The lists change; the established names recur. A plausible reading is that these brands are corroborated across so many independent sources, from reviews to retailers to news coverage, that any reasonable sample of the web includes them. The studies cited here do not test that explanation directly, but it fits both the SparkToro pattern and the Authoritas contrast between real and fake experts.

Signals inside the content

How a page is written also affects how much of it a generative system uses. The GEO paper, presented at KDD 2024, found that adding citations from reliable sources, credible quotations and relevant statistics improved a source’s visibility in generative answers by 30% to 40% on one of its measures, in the authors’ test setting. Rewriting text in a more persuasive, authoritative tone did not produce significant gains. Evidence helped; the appearance of authority did not.

Trust is not the same as accuracy

Appearing often does not mean being described correctly. In the Authoritas experiment, when models were asked directly about the fake names, they treated them as credible authorities 29.5% of the time, sometimes inventing career histories. And in a study of references generated from model knowledge alone, published in the Findings of NAACL 2025, Algaba and colleagues found models favored already highly cited papers and also produced non-existent references resembling real ones.

Models can be confidently wrong about sources, and they can favor what is already prominent. A trustworthy source, in the sense that matters to a publisher, is one that is both included and represented accurately.

What this means when you assess a source, including your own

1. Separate inclusion from accuracy. A source can appear often and still be described wrongly, as the Authoritas and Algaba findings show. Judge the two separately.
2. Look for corroboration the source does not control. Independent confirmation was the clearest difference between the real and fake experts. If the only sources describing an organization are its own, the Authoritas contrast suggests it will carry less weight.
3. Check how claims are supported. In the GEO experiments, cited sources, quotations and statistics helped where an authoritative tone did not. A page that shows its evidence gives a system more to use than one that only asserts.
4. Know which layer you are looking at. For Google’s AI features, trust runs through Google’s ranking systems. For training, it runs through how data was weighted before a model was released.

Trust, as these systems reflect it, is not something a source can declare. It is visible from outside: in independent sources confirming who a source is, in consistent facts across the web, and in pages that show their evidence. There is no published trust score to aim at. What models tend to reflect is the web’s consensus about you, and that consensus is built from sources other than your own.