Citation Rate (AI Citation Rate).

The citation rate answers a simple question: out of a hundred answers that ChatGPT, Perplexity, Gemini or Google AI Mode give to a defined set of questions, in how many does my domain appear as a source? The figure is always bound to that question set and is meaningless without it. Two variants are in common use: the linked citation, where the domain is listed as a source, and the plain mention in the answer text without a link. Record them separately, because they arise differently and behave differently over time.

Measuring is either manual work or tool work, never an accident. First a fixed set of twenty to fifty questions is defined, chosen to reflect real purchasing decisions rather than vanity topics. Then each question is asked several times in each assistant, because the answers are probabilistic and vary between runs. Log the date, the model, the exact wording, whether the domain appeared and in which passage. A single run is not a measurement, it is an anecdote. Only repetition across several weeks shows whether a change has actually taken hold.

The citation rate behaves quite differently from a ranking position. It jumps because models are updated, because retrieval pulls different sources, or because a third-party page adds a new mention. It is also strongly topic-dependent: a provider can be named regularly for questions about GEO consulting in Zurich and never at all for questions about website costs. For that reason it should be reported per topic cluster rather than as one house number, and read as progress only against your own previous month and against competitors in the same set.

A case from consulting practice: a provider wanted to know whether a new advice section was having any effect. Before publication, a set of thirty questions was run three times each across three assistants and the hit rate recorded. Eight weeks later the identical set was run again, unchanged. Only that before-and-after comparison allowed a defensible statement, precisely because the set had stayed constant. Anyone who adjusts the question set in between is no longer measuring progress — only a different set.