AI citations move more than most people checking once ever notice

Three separate mechanisms drive citation change, and they stack rather than substitute for each other. The model itself is probabilistic: it is built to predict a plausible next step rather than retrieve one fixed answer, so some variation between runs is built in by design.

On top of that, the companies running these systems periodically retrain or fine-tune the underlying model, which can shift which sources it trusts and how it frames an answer without any change on the cited page. And systems built on retrieval-augmented generation re-search the web on each query, so a change in which pages get pulled into that search changes the citation set even when the top-ranking pages themselves are unchanged.

Volatility at the citation level does not mean the underlying work of being legible to an AI system is wasted. It means the unit of measurement is wrong if it is a single check. A business that gets cited once and stops looking has learned almost nothing about whether it is actually positioned well, because the same query run again next week could return a different picture in either direction.

The more useful habit is to check the same handful of queries on a recurring cadence and read the trend rather than any single result, the same discipline multi-engine AEO already asks for across platforms, applied across time instead. What stays comparatively stable underneath the churn is citation share measured over a window, and the structural clarity, entity, category, claims, that gives a system something safe to cite in the first place, whichever exact week it happens to look.