Method
How Upcite measures
A number is only worth something if you know where it comes from. That is why this page explains exactly how Upcite measures. Last updated on 12 October 2026.
1. We ask without your client's name
A measurement question never names the company we measure. We ask the question the way a customer would ask it, for example: which [service provider] in [region] do you recommend? This way we measure whether a language model mentions your client on its own, not whether it recognises the name.
2. The questions come from search data
Measurement questions match what people in your client's market search for. Upcite suggests measurement questions based on keyword data. A good measurement question describes a real choice, with a need, a situation and the criteria someone uses to choose. A question never steers towards your own client.
In the keyword module we show search volume as underlying search potential, with source, country and reference date. It does not predict how many people ask a language model the same question.
3. Several language models, without a personal account
We measure in ChatGPT, Gemini, Perplexity, Google AI Overviews and Microsoft Copilot four times a month, and in Claude once a month. We measure without a personal account, so an answer is not tailored to earlier conversations of the person measuring.
4. Every question several times
A language model does not always give the same answer to the same question. That is why we ask every question four times a month, in separate measurement rounds. How much the answers differ between rounds is shown as volatility.
5. How we assess an answer
| Outcome | Meaning |
|---|---|
| Mentioned first to fourth | Your client appears in the answer as the first, second, third or fourth company |
| Mentioned outside the top four | Your client is mentioned, but after four other companies |
| Mentioned without ranking | Your client is mentioned in an answer without an order, such as a simple list |
| Not mentioned | Your client does not appear in the answer |
The position shows where your client appears in the answer. It is not a ranking the language model assigns itself. For every answer we also assess the sentiment and which sources the language model cites.
6. Numbers are counts
Every percentage is an exact count within the measurement. Being mentioned in 18 of 60 answers equals 30.0%. The measurement card shows how many answers were counted and how many are missing. We do not convert measurement results into visitors or revenue.
7. Volatility
Per question and per language model you see how often your client is mentioned, how the position varies and whether the cited sources change. We distinguish between differences within one measurement round and differences between measurement days. We do not use an unsubstantiated threshold for 'stable'.
8. Facts about your client
A language model can say something incorrect about your client, such as a wrong address or a service the company does not offer. With Fact check you rate every factual claim as correct, incorrect, outdated, about the wrong company, or not supported by the cited source.
No source found does not mean the language model is making something up. It means the claim cannot be verified.
9. The measurement card
Every report comes with a measurement card. It shows:
- the topics and the number of measurement questions
- language and region
- the AI systems measured, with model version where visible
- the measurement days and the number of repetitions
- how many answers are planned, assessed, still open or missing
10. What we do not promise
Upcite measures visibility in language models. It does not guarantee a mention. Language models change, and every measurement applies within the limits shown on the measurement card.