There is nothing like a good old fashioned survey. If you want to understand what someone thinks, how they feel, or who they are, why not just ask them?
I know that algorithms and modelling can help predict who someone is or what their preferences are. I know that we do not always have the opportunity to survey people, to survey enough people (for statistical significance), or to survey the right people (representative sample). Sometimes other data collection or data aggregation methods can be easier, cheaper, or more efficient. I also realize that survey responses might be less than accurate or even biased at times.
However, I do think that there are situations where survey or poll questions can be very effective. One example would be to use surveys to understand basic user behavior and usability of a new website or mobile/tablet app. You could certainly look at and interpret analytics, clickstream, or user flow data. Surveys though, can give an even better sense of the "whys" behind the behaviors. You could additionally get valuable information on sentiment, expectations and future intent. For example, Condenast used surveys to understand who their tablet digital issue readers were (3/4 were already subscribers or newsstand readers of print issues). They also gained some info on intent, 4 out of 5 were likely to download the next month's issue. Other information on sentiment could help drive future content.
Another good use of even a single poll question is to help ad networks, exchanges, and other similar data driven companies improve their targeting or optimization models. A single piece of audience information as basic as demographic data or as important as in-market buying potential could drastically improve an ad targeting model and improve ad response or effectiveness. Predicting whether someone is in market for an appliance is one thing, actually asking someone when they are planning to make a purchase next seems a little more definitive to me.
In the end, a combination of surveys and other methods of data collection, modelling, and analyses put together will be needed. But don't under estimate the power of a question.