The date a trend actually broke
Change Point Detection
Finds the day a metric changed level rather than drifted, so a site change or an algorithm update can be dated instead of guessed at.
How it works
Each candidate day is tested by comparing the mean of the window before it with the mean of the window after, and the day with the strongest separation relative to the surrounding variance is reported as a change point. This answers a different question from anomaly detection: an anomaly is one odd day, a change point is the day after which the normal level is different.
What you get
- Detected change points with their date, direction and magnitude
- Before and after averages for each one, so the size of the shift is concrete
- Overlay of known Google updates, to separate your change from theirs
- A defensible date to align with a deploy, a migration or a content push
What it needs
At least 14 days on each side of the suspected change.
What it will not tell you
A detected date is a statistical best fit, not a cause. Gradual decline has no change point to find, which is what Time-Series Decomposition is for.
Runs on demand
Experimental analyses are not part of the nightly job. You open the report, set the period and run it, so the cost lands only when you want the answer.
The rest of Experimental Analytics
Ready to see what your Search Console data is hiding?
Connect a property and every report fills in immediately. No credit card needed.