1. Aggregate complete weeks
Daily search data carries day-of-week noise. We aggregate Monday through Sunday, treating omitted zero-click days as zero and dropping partial weeks at both edges. A mid-week export cannot masquerade as a decline.
2. Build a robust baseline
The eight weeks before the four-week test window form the baseline. We use the median rather than the mean, then median absolute deviation to estimate normal noise. One viral week cannot redefine normal performance.
3. Require sustained decay
A page is marked DECAYING only when at least three of the four recent weeks fall below both the noise floor and a 25% relative-drop threshold. Severity is the estimated weekly clicks lost: baseline median minus recent mean.
4. Detect “flat, not down”
Theil–Sen uses the median of pairwise slopes. It is less sensitive to outliers than ordinary linear regression. A page can be FLATLINED when it previously grew materially, then settles into a near-zero recent slope.
What this cannot know
The engine detects a shape. It does not prove a cause.
Seasonality, migrations, tracking changes, SERP redesigns, cannibalization, brand events, and changed search intent can produce similar trajectories. The evidence narrows attention; an agency strategist still reviews the diagnosis before work reaches a client.
Minimum acceptance standard
The pre-committed dogfood test is fewer than 20% false positives in the top-ten severity queue on a real owned property. If the top of the queue wastes human attention, the engine is not ready to automate delivery.