Almost every discussion about AI Search seems to revolve around one word:
Trust.
AI trusts websites.
AI trusts brands.
AI trusts experts.
Become a source AI trusts.
Increase AI’s trust in your business.
But what exactly does trust mean in these statements? Are people always talking about the same thing—or has a familiar human word quietly become shorthand for several very different technical processes?
Are We Really Talking About Trust?
When people say that AI “trusts” a website, they may be referring to very different things. Sometimes they mean the reputation of a brand or the authority of an expert. Sometimes they mean the factual accuracy of the information. Sometimes they mean whether that information is corroborated by other independent sources.
And sometimes we simply observe the outcome—a website appears in an AI-generated answer, is cited, or is recommended—and explain the result by saying that AI “trusted” it.
But an authoritative source does not necessarily provide the best answer to a particular question. A retrieved document may never influence the final response, and information that is used may not be attributed to its original source.
So while saying that “AI trusts this website” is easy to understand, the phrase alone tells us very little about what actually happened.
Can Trust Be Measured?
The idea of computational trust existed long before AI Search.
For many years, researchers have explored ways to evaluate the reliability of information, the reputation of participants, and the trustworthiness of different sources.
One well-known example is Google’s 2015 paper, Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources. Instead of estimating a website’s quality based on popularity or the number of incoming links, the authors proposed evaluating a source according to the estimated accuracy of the factual information it contains. The resulting metric was called the Knowledge-Based Trust score.
This example shows that trust can indeed be treated as a technical, measurable concept. However, its meaning here is very specific. It refers to the factual accuracy of a source—not to a general “relationship” between an AI system and a website.
Technical research also tends to specify exactly what is being evaluated. For example, Retrieval-Augmented Generation with Estimation of Source Reliability estimates source reliability by comparing information across multiple sources and combines that estimate with relevance during document selection. Not All Contexts Are Equal: Teaching LLMs Credibility-aware Generation focuses instead on the credibility of the retrieved context and how language models should use evidence of different quality when generating responses.
In other words, technical research does not necessarily avoid the concept of trust. But behind that single word usually lie much more specific properties, measurements and operations.
One Word, Many Different Processes
When people say that AI “trusts” a source, the word trust often serves as an umbrella term for several different characteristics, system decisions and observable outcomes.
- Source authority, credibility, reputation — How authoritative and trustworthy a website, organisation, brand or expert is considered to be.
- Factual accuracy, source reliability — How accurate and reliable the information within that source is considered to be.
- Corroboration, cross-source consistency — Whether the information is confirmed by other independent sources.
- Retrieval, source selection — Whether the document is retrieved and included among the materials available for generating an answer.
- Ranking, relevance scoring — How well that document matches a particular query compared with other available documents.
- Evidence weighting — How much weight the system gives to information obtained from that source.
- Source use, attribution, citation — Whether the information is actually used in the response and whether the source receives explicit attribution.
What this list actually mixes together are properties of the source itself, choices the system makes, and things we observe after the fact — three different kinds of thing hiding under one label.
These factors are often related, but they are not the same thing.
Authority may influence whether a source is selected or how much weight it receives, but a document may be chosen primarily because it is the most relevant to a particular query. Accurate information may be used without being cited, while citing one page does not necessarily imply unconditional trust in the entire website.
Even “Trust” Doesn’t Guarantee Inclusion
Even if a system considers a source to be trustworthy, that alone does not guarantee that it will appear in a particular answer.
The final decision may also depend on relevance to the user’s query, the completeness of the available information, the need to corroborate factual claims, the diversity of sources and viewpoints, model limitations, and the instructions given by the system’s developers. The broader conversation, user preferences, location and personal context may also influence the outcome.
In other words, the perceived trustworthiness of a source is only one factor among many—not a guarantee of citation, mention or recommendation.
So What Should We Do with Trust?
There is nothing wrong with using the word trust as convenient shorthand.
The problem begins when it becomes the objective itself: “We need to increase AI’s trust in our website.”
That sounds appealing, but it skips both the diagnosis and the plan for fixing it.
The first step is to identify where the actual issue lies.
The first thing to check is simple access: can the system even retrieve the document? From there the questions get more specific — does it actually answer the query, and is its information corroborated elsewhere? Further along the chain, there’s the matter of weight: how much does the system trust what it found, does that information make it into the final answer, and is the source ever named?
From the outside, the outcome may always look the same: the website did not appear in the answer.
But the underlying causes—and therefore the required work—may be completely different. The issue could involve technical accessibility, relevance and completeness of the content, factual accuracy, external corroboration, source reputation, or the way attribution is handled.
That kind of diagnosis determines not only what needs to be improved, but also what success should actually look like. Without it, promising to “increase AI trust” explains neither what will change nor how the outcome should be measured.
That is perhaps the most practical conclusion of all.
In AI Search, it is far more useful to ask which specific property or process needs improvement than to talk about improving trust in the abstract.
Only then do the necessary actions become clear, the expected outcomes become measurable, and meaningful progress becomes possible.
