It matters. Anything except a truly random sample of the population (essentially impossible) invalidates the statistical guarantees of polling. Consequently, polling involves a large number of processes to both get the most representative sample possible, then adjusts the results to reflect the biases they are aware of. This is something of an artistic process, and is one reason why pollsters have house effects/biases: they each do things slightly differently.

Having said that, the general feeling is that online polls are as reliable as offline polls nowadays after these adjustments have been made, partly because polling by telephone has a well-established set of biases associated with it too.