Stratify before you split
Control and test buckets must start with matched traffic, templates and intent, or the result is noise.
One variable per test
Titles, internal links, content blocks and schema get tested separately. Combined tests cannot be read.
Let it run through a crawl cycle
Search engines need time to recrawl and re-rank. Reading a test early is the most common mistake.
Read clicks and impressions with intervals
Search Console data, confidence intervals and a pre-registered decision rule. No eyeballing charts.
Revert losers fast
A losing test is reverted the day it is called. Rankings lost to a bad template are expensive to win back.
Roll winners out as a release
Winners ship site-wide with the same QA as any product change, and the test log records why.
What we deliver for URL bucket testing
The concrete work behind these principles, on the Services page.