Build the prompt set first
Twenty to forty prompts in real buyer language: category questions, comparison questions, 'best tool for X' questions, and brand questions. Freeze the wording - changing prompts between runs destroys comparability.
Log four things per run
Whether your domain is cited, its ordinal position among citations, which competitors appear, and the sentiment of any sentence naming your brand. Those four fields support every chart worth building.
Convert to a share-of-voice metric
Citation share is your cited prompts divided by total prompts; competitive share is your citations divided by all cited domains. Report both - the first tracks coverage, the second tracks dominance.
Close the loop with content
Every uncited prompt is a content gap with a name. Either the passage that would answer it does not exist, or it exists and is not quotable. Fix in that order and re-run.
Correlate with analytics
Segment referral traffic from AI hosts and compare it against citation share. A rising citation share with flat referrals still delivers brand exposure inside the answer, which shows up later as direct and branded search.