The 16-of-32 to 12-of-32 example needs a within-week baseline before I'd call it a real loss. Those 32 answers combine different prompts and engines, and one generation per cell can move quite a bit without anything changing on the site.
I'd repeat each fixed prompt a few times in fresh sessions, retain the answers and citations, then compare mention frequency per engine and prompt. Record search mode, locale and the model/version when exposed. Keep “recommended by name”, “linked as a source” and “described accurately” as separate columns; a brand can lose a citation without disappearing from the recommendation.
Also, “why didn't you include us?” is a follow-up generated explanation, not an execution trace of the original answer. It can suggest things to verify, but shouldn't establish the cause of the drop.
For a team building its own monitoring pipeline, QuanticData's AI Visibility API is worth evaluating alongside a manual baseline; I work on QuanticData. I'd still validate automated observations against the particular consumer/search experience the team actually cares about, rather than treating every API answer as identical to that UI.