This study offers evidence of the existence of switching costs on the Internet. It uses more flexible methods than previously possible to separate switching costs from serially correlated unobservables at Internet portals. The data contain nearly 1,000 observations per household, allowing for household-specific regressions that control for all household-specific heterogeneity. The results show that households exhibit switching costs. The loyalty generated by these costs drives a large fraction of portal visits and generates considerable revenues; however, these revenues are not large enough to justify the losses incurred by Internet portals in the 1990s while building market share. The results also suggest that random coefficients models overestimate true state dependence.