The situation
Twelve sites, all hit by the Helpful Content System’s sitewide, multiplicative downranking. The industry folklore said “wait and pray.” We instrumented instead.
The approach
We treated the classifier as an object to be measured, not a curse to be lifted. Each site was tracked through the resample cycle with a fixed remediation applied at a known date, so the recovery — when it came — could be attributed.
- Sitewide quality floor. The classifier scores the site, so thin pages anywhere drag everything. We pruned, not padded.
- Resample timing. Remediation landed early in the cycle so the next resample would read the improved corpus.
- Holdout. Two sites were left unremediated as controls.
The result
The remediated cohort recovered on a 60–90 day cadence with a +148% median rebound; the controls did not move. The resample curve is not mystical — it is a measurable cadence you can plan a remediation against.
The lesson
Most “algorithm recovery” is guesswork dressed as expertise. With a holdout and a calendar, it becomes engineering. Measure the cadence, time the fix, attribute the result.
