Where AI finds the evidence about your brand

Steven Wilson-Beales looked at 4,003 AI answers about news publishers. In 76% of them, no about page was among the sources. The evidence depended on the question: about pages for who a publisher is, pricing pages for what it is worth, and pages about standards and corrections for whether it can be trusted.

We wanted to know if that holds outside the news. So we asked the same kind of questions about banks, shops, airlines, telecom companies and more.

25 well-known Dutch brands in eight sectors. For each brand, 12 questions in Dutch: three for each of four claims a brand makes.

We asked every question twice, to ChatGPT and to Gemini. Once without web search, so the answer comes from what the model learned in training. And once with web search, so the model can look things up and cite its sources. That makes 1,200 answers.

Then we sorted every cited source. Is it a page on the brand’s own site, or somewhere else? And what kind of page is it: an about page, a price page, a complaints page, a review site, Wikipedia, the news? Two language models did the sorting, and a third decided where they disagreed. An about page here means any page about the organisation itself. Whether it is a good one is a different question.

Last, we put each answer from memory next to the answer with search to the same question, and marked whether they say the same thing.

1. Your own site is the evidence AI uses most. When AI looks something up about you, it usually lands on your own site: for who you are, for what you cost and for what you know. That makes your site the place where you can put things right. And the starting point is the page that says clearly who you are: the True Root.
2. One page is not enough. The about page carries a quarter of the evidence for who you are, and almost nothing for trust or price. Every claim needs its own proof: complaints and service pages and clear policies for trust, clear offer and price pages for value, pages about your people and your work for expertise. Linked from the root, and telling the same story. That is the evidence architecture.
3. Keep the facts only you can set current, everywhere. Memory is behind on who leads and owns a company, and Gemini often does not look it up. So the right facts have to be on your site and in the places that feed the models: Wikipedia, registers, business information sites.
4. Trust is decided elsewhere, but you can feed it. A third of the trust evidence is review sites. Your own complaints and service pages are the second largest source. Publish how you handle complaints, and the model has something of yours to cite next to the reviews.

The source grader turns this into a checklist of 67 sources, in the order that works: your own site first, then the copies of it, then what others write.

• 25 brands, on one day (2 October 2026), with questions in Dutch. Another moment or other brands can give other numbers.
• Two models: ChatGPT and Gemini. Gemini searched for only half of the questions, so it has fewer sources.
• Language models sorted the sources and compared the answers, two of them, with a third where they disagreed. That is consistent, not flawless.
• A difference between memory and search is not proof of an error. We did not judge which answer is true.
• The research shows which pages AI cites. It does not show that adding a page leads to more citations.