Without a baseline you cannot say whether a change to your website worked: there is nothing to compare with. So it is done before touching anything and kept in full, with every observation, not only the summary.
For the comparison to hold, the next measurement repeats the same questions, on the same platforms, with the same rule for errors. If you change the set of questions, you start a new baseline.
A simple baseline can be done by hand in twenty minutes: six questions on two platforms, twelve observations.