Because AI answers vary, share of voice is measured by sampling: run your core questions through the engines repeatedly and track the proportion in which you show up.
Some tools also weigh sentiment, not just frequency, since a negative mention is not the same as a recommendation. It is directional, not exact: Trakkr's March 2026 eight-model study found AI tools agreed on a brand's mentions only about 43.9% of the time, so the absolute number shifts with the engine. The trend across repeated samples is what tells you whether your GEO work is moving you into more answers over time.