Compete is often wildly off. Compare the different services that offer similar analytics and you'll see how often they disagree.
The rise of the official and unofficial non-browser clients have a hand in it I imagine.
Or those numbers could just be bollocks.
Why do people still think compete, quantcast, alexa, etc are accurate when they don't have tracking pixels installed on the sites they're showing traffic for?
There was an interesting comment recently on this topic, someone explained that some of the larger tracking companies purchase traffic logs from American ISPs and although their traffic reports aren't necessarily accurate (because they don't have all the data -- as shown by the github employee below) it's not complete guess work and they can be considered representative for sites used by the average internet user.