They do this by aggregating the results from multiple polls to level out the bias and produce a more accurate picture of the field overall. We have the same problem with programming language surveys.