I was using cohorts to define a type of user based on a user property being one of several values (call this Cohort A). I tried to then look at the negative of that cohort using the same set of values but changing property “is” to property “is not” (call this Cohort B). When I did this, it seemed that both cohorts (Cohort A + Cohort B) do not add up to the total number of users. In fact, there is a huge discrepancy and even when I tried to create a third cohort that represents Cohort A or Cohort B (call it Cohort C), that too does not add up to the total users for that time period.
However, if I use the same list of user property values within a segment and create 3 segments in a chart following the way I had defined the cohorts, I do find that Segment A + Segment B = total users. and that Segment C (which can only be Segment A and Segment B) is zero as expected.
All this said, the segments work as I would expect, but cohorts seem to be inexplicably inaccurate.