Read the rest of the comments here. Everyone else is having similar problems. Also, please spare me the expected "This is all anecdotal/sample size of 30" follow-up. Perhaps you can answer the real question -- is this valuable to anyone if it has a significant error margin? I think it wouldn't be allowed a margin of more than 1% if it were to be useful, and even that is kind of pushing it if you intend to do anything important with the data -- if 20 sites are tested per visitor, a 1% error rate would mean that an incorrect detection would occur every fifth visitor or so. That's enough to allow plausible deniability in my book.

FWIW, I added several improvements, and according to the built-in survey, it works for about 95% of all visitors. If you had bad results initially, clear you cache and give it a second try.