I've read somewhere (I forget the source, sorry - I think the MS Office developer's blog?), that when you do a survey of users asking them about what features they would like to see in your software/website, they will more often than not say that they want every little thing, whereas collected metrics show that in the end, most people don't use 99% of these features. The general message from the blog post was that you shouldn't ask people what they use, you should track it for yourself.
This leads to an unfortunate chicken-and-egg situation when trying to figure out what new feature to add next. Without the feature already in place, I can't measure how much it's actually being used. With finite (and severely stretched) resources, I also can't afford to add all the features and then remove the unused ones.
How do you find out what will be useful to your users? If a survey is the only option, do you have to structure your questions in certain ways (eg: don't show a list of possible features, since that would be leading them on)?