http://www.sciencebuddies.org/science-fair-projects/project_ideas/Soc_participants.shtml
Basically, 100*n^(-1/2), with n being number of surveys returned, would be percent error of the survey. If I can get 150 people, I will have about 8% error on the questions that are multiple choice or yes/no answers. I think that is a pretty good accuracy for a managable number of surveys.