We spent a quarter asking the same thirty prompts and writing down whether we were cited, linked, paraphrased, or absent. It is a terrible pipeline. It is also more honest than a vendor score that will not show its work.
Patterns that survived the tedium: pages with a direct answer in the opening still got mentioned more often. Pages that buried the answer under a table of contents got summarized from a competitor. Dates on the page helped when the prompt asked for something current; they did nothing for evergreen comparisons.
The point of counting by hand is not to scale it. It is to know what “AI visibility” means before you automate it. Then you put those dates next to the edits you made, the same as you would for a ranking.