The article analyses how different methods of recruiting Internet users for online surveys affect survey results in terms of marginal distributions and associations between variables. The general hypothesis is that self-selection processes in some kinds of online surveys, especially in open online surveys, bias marginal distributions: those more interested and involved in the topic (in our case: politics) are expected to be over-represented. This bias furthermore causes (political) attitudes to be much more structured among participants of (open) online surveys. The hypothesis is tested using data from face to face interviews with Internet users, an access panel of online users and an open online poll, all collected in the context of the 2002 German federal election. Empirical analyses show that both marginal distributions and associations between variables stemming from the open online survey differ clearly from those based on the other surveys. Additionally, the analysis reveals that standard weighting procedures do not reduce these biases substantially.