Philipp Lenssen at Google Blogoscoped posted yesterday about how the accuracy of Alexa data is really poor. The specifics of his post are really a spoof as all the "data" is made up. Philipp is in fact spoofing the Alexa methodology itself as he indicated that the data: "uses gut feeling from a selected sample group (me) as data source". Alexa's accuracy problems are well known.
The underlying problem is that the Alexa data is derived from users who use the Alexa toolbar. At the end of the day, the audience is just not large enough, and the dependency on a willingness to install the toolbar introduces a natural bias into the date. My own experience suggests that these problem become worse and worse as you deal with lower and lower traffic level sites. Have a site that gets 20,000 visitors per day? You are not really on the map with Alexa at that level.
However, I still use Alexa as a tool. It was a blog post by Avinash Kaushik that taught me how to still use it as an effective tool. Quite simply, use the Alexa feature that shows comparative traffic levels to compare your site's traffic to that of your competitors.
Because your competitors are in the same business as you are, the bias problem no longer is a factor to worry about (because the bias will affect all the compares sites equally). For most businesses this will provide a quick way to compare the relative web site traffic levels in their industry. So the accuracy problems are real, but there is still a way to use the tool to extract useful information.