How ChatGPT picks its sources

In September 2025, we asked ChatGPT 400 questions about topics where our clients could have been cited: Geneva watchmaking, Swiss web agencies, private banking, Michelin-starred restaurants. For each answer, we listed the cited sources and looked at what they had in common. What came out was surprising.

What barely matters

First, what we thought counted — and barely does: PageRank, Google position, number of backlinks, domain age. Across 400 answers, the top-cited site was almost never the first Google result. Sometimes it was on page 3.

"The site cited first by ChatGPT was almost never the top Google result. Sometimes it was on page 3."

The three signals that count

Cross-referencing the 400 answers, three factors emerged consistently. We ranked them in order of empirical importance:

1. Explicit named entities. The site clearly states who it is (brand name, founders, year, location) in the HTML — headings, subheadings, JSON-LD. Without that, ChatGPT cannot identify you.
2. Complete Schema.org markup. Cited sites had 2.3× more JSON-LD blocks than average. Organization, LocalBusiness, FAQPage. The correlation is nearly linear.
3. A declarative tone, not a marketing one. "We are a web agency in Geneva founded in 2020" gets cited. "We bring your digital ambitions to life" never does. ChatGPT looks for facts, not slogans.

The takeaway

To get cited by ChatGPT, you need to write like an encyclopedia, not a brochure. The sites that come out on top are the ones that state plainly, in clean HTML, who they are and what they do.

That is exactly what Swissalytics audits. Paste your URL: we tell you, in 30 seconds, whether an LLM can clearly identify you or not.