The failure mode of rank alerts is not missing drops. It is the sixth false alarm in a week, after which every alert, including the real one next month, gets archived on sight.

Alerting is threshold design. Here is the ruleset that fires on action-worthy moves, stays silent through ordinary wobble, and keeps its owner's trust for years.

The wolf cycle

tight thresholds alert on every wobble daily false alarms nothing needed doing alerts ignored archived on sight real drop read weeks late The system failed at step one: thresholds tuned to wobble guarantee everything after.
Alert fatigue is designed in, at threshold-setting time. So is its absence.

The four principles

Scale thresholds to the position band. The normal wobble bands are the floor: an alert threshold below the band's natural dance is a false-alarm machine. Positions 1 to 3 justify sensitivity; positions 11 to 30 need a much wider trigger.

Require the move to sustain. A drop that alerts on day one alerts on noise. Three consecutive days of holding the new level filters the experiments and re-shuffles out, and costs little: real problems are still there on Thursday.

Trigger on scope, not just size. Twenty keywords each slipping two positions is a bigger event than one keyword dropping ten, and size-only thresholds miss it entirely. A same-direction move across a meaningful share of tracked keywords deserves its own, immediate trigger.

Weight by value. The keywords mapped to money pages get tighter rules than the long-tail flock. Tiering is what lets the important alerts stay sensitive without the whole system going noisy.

The ruleset

TierAlert whenDelivery
Money keywords3+ positions, sustained 3 daysDigest, top of it
Page-one exit (money)Any money keyword leaves page oneInstant
General tracked set5+ positions, sustained 5 daysDigest
Sitewide pattern~20% of keywords moving one direction in a weekInstant
Gains into striking distanceAny keyword entering positions 5-15Digest, as opportunities

Two instant triggers, everything else in a daily or twice-weekly digest. The page-one exit rule exists because the traffic cliff between positions 10 and 11 makes that single step worth a same-day look; the sitewide rule exists because patterns mean something structural or an algorithm update, and both reward early forensics.

The gains row is the one people skip and should not: striking-distance arrivals are the quick-win queue refilling itself, and an alert is how they get acted on while fresh.

The digest that gets read

Format decides survival. A digest that lists forty movements gets skimmed to death; one that answers "what needs me today?" gets read for years.

Three sections, strict: needs action (threshold-crossing moves, each with the affected page), worth knowing (sustained trends forming), and opportunities (the striking-distance arrivals). Anything that fits none of those sections stays in the tool, on the graphs, where curiosity can find it voluntarily.

The monthly tuning loop

Once a month, count two numbers: alerts fired, and alerts that led to action. Below roughly half acted-on, the thresholds are too tight somewhere, and the offenders are usually visible in the log: the same volatile keyword tripping its rule weekly earns itself a wider band or an exit from the money tier.

Alerting is a garden, not a monument. Ten minutes of monthly pruning is what keeps the system trusted, and trust is the entire product: an alert that arrives believed is worth twenty that arrive filtered.

The one-line takeaway: alerts fail by firing, not by missing. Scale thresholds to the position band, require moves to sustain, trigger instantly only on page-one exits and sitewide patterns, digest the rest in three sections, and prune monthly until most alerts lead to action.