For the top 1000 or so it's probably pretty accurate.
Why not?
The system they use for metrics is biased. They track the visits of the users who have installed the Alexa toolbar in their browsers. It's really hard to get a decent sample of Internet users using that data collecting method.
Alexa now does direct tracking, in a somewhat similar manner to Quantcast. I don't think very much of their data is yet represented in that manner however.