The end date is decided before the start. Stopping when a variant looks good produces a number that will not repeat, so recording the planned duration and sample in advance is what makes a result defensible.
Hypotheses have to be written down first. A prediction recorded before the data arrives prevents the retrospective explanation that makes every outcome sound intentional, and clients notice the difference.
The archive is the compounding asset. Three years of tests, including everything that failed, is what lets a practice tell a new client what usually works in their sector without charging them to rediscover it.
Sites move while tests run. A promotion launched or a template changed mid-experiment invalidates the result, so a log of what else happened on the site is part of the evidence rather than an aside.