This paper considers the implications of new analytic advances on Internet-based surveys of non-random samples. The empirical evidence presented indicate that Internet-based forecasts in the 2000 U.S. elections were two times more accurate than telephone forecasts in common races, and that propensity score adjustment, a technique designed to minimize biases associated with non-random samples, was a main reason for the difference. As the credibility of any predictive research approach may hinge on its ability to pass this acid test, the evidence is important, with implications on research methods other than Internet-based surveys. For instance, the technique of propensity score adjustment could also be exploited by research organizations with panels measuring Internet activity, off-line purchase behavior, and TV viewing behavior. Such organizations could depend on propensity score adjustment to link attitudinal, opinion and other information to behavioral information collected through these panels without compromising data quality or substantially increasing the cost of data collection.