This paper firstly identifies the primary reasons for the growth of online research and then reviews the principal concerns many traditional researchers have about the quality of data collected and summarises some of the findings from various trials conducted to evaluate those concerns. The author points out that, throughout the literature there are relatively few references to the question of the probability with which a response has been obtained, and the fact that a self-selected sample, inherent to online research, is not compatible with the known probabilities of response (often equal probabilities) that underlies traditional sample theory and the interpretation of results. Inter alia, it is then claimed that online work must always involve a potential for bias that is largely unknown. Moreover it is asserted, that the growing acceptance of Bayesian methods and associated weighting procedures to adjust for changes in relative probabilities are totally inadequate to deal with the variations in probability encountered with online data. A discussion of the use of mixing modes for data collection (primarily telephone landline, telephone mobile and online) is used to illustrate the difficulties of interpreting the relative probabilities and their effect on the data. Finally, the paper examines some of the recent changes introduced in online sampling and concludes that the results do not alter the fundamental problem of unknown ‘selection’ probabilities and associated unknown biases in results. The conclusion drawn is that the ‘best bias’ may be cheap, but that the consequences may be very expensive!