People have done a lot of experimentation with this. The number goes up when someone joins, and goes down when they cancel, by the correct amount. People watched from early on, and the counter started at or close to zero, grew continually for a few years, them tapered off to a level amount, then took a dip for a while, then flattened again. There is no way to opt into or out of it on your own. It seems pretty reliable. What's your evidence that it is not reliable?
I've seen the source for the "adding/removing a user" test, and for the totals, but not for the "from the beginning" or the "took a dip" data - have a source for those?