During the last decade, Web and IVR (Interactive Voice Response) polls have spread in many countries and particularly in small markets where media outlets do not have the resources necessary to sponsor electoral polls conducted using “classical” modes. In addition, there is no incentive to put money in classical telephone polls if IVR or Web polls perform well and cost much less. But is it really the case? In this presentation, I analyze the performance of Web and IVR polls in various contexts, namely recent Canadian elections at the federal and provincial levels, the Scottish referendum of 2014 and the US election of 2012. I examine poll performance according to mode of administration on two aspects: a)systematic error, i.e., the capacity to forecast election results without bias, and b) random error, i.e., the level of variability in estimates, taking into account the sample size. I show that Web and IVR polls tend to have systematic bias that may – and should – be corrected in order not to mislead voters. IVR polls also tend to show substantial random error in specific contexts. In addition, bias is not always present so that we need to be able to identify which context and which specific methodological features are more likely to lead to bias. I examine these questions and conclude on how to correct bias, if possible, and, on the necessity to inform voters about the possible biases of the polls they are consulting.