We rebuild the page that decides the sale, then run it against your current page on live traffic until revenue per visitor gives a clear answer.
Usually the product page, because that is where the money is decided. Rebuilt from the ground up rather than patched with a few tweaks.
Your visitors split between the new page and your current one, running at the same time. Seasonality and traffic mix drop out of the comparison.
Two or three tests a month after the first win, each one measured the same way. Fewer opinions, more tested changes.
Every figure we publish was measured in an A/B test on the brand's own traffic, not modelled or estimated.