For example, a very big site, you have enough trackees to make a reasonably statistically valid sample. But smaller sites, you don't see as many trackees, so you set the error bars higher.
Sure, I agree that the system is flawed anyway, seeing as how I don't know a single person who installs the alexa toolbar, and amongst techie sites it's usage is probably 0.