Web surveys are frequently based on samples drawn from panels with large amounts of nonresponse or haphazard selection. The availability of large-scale consumer and voter databases provides large amounts of auxilliary information for both panelists and population members. Sample matching, where a conventional random sample is selected from a population frame and the closest matching respondent from the panel is selected for interviewing, is proposed. It is shown that under suitable assumptions (primarily ignorability of panel membership conditional upon the matching variables), the resulting survey estimates are consistent with an asymptotic normal distribution. Simulation results show that the matched sample estimators are superior to weighting a random subsample from the panel and have a similar sampling distribution to simple random sampling from the population. In an example involving the 2006 U.S. Congressional elections, estimates using sample matching from an opt-in Web panel outperformed estimates based on phone interviews with RDD samples.