Active video games (AVG) have become widespread as more physical interfaces are introduced in video games. Lab based studies have indicated that AVGs can increase the amount and intensity of physical activity compared to non-active games and TV, however, the long-term effectiveness of AVGs has yet been established. In fact, most of the existing studies show a reduction of interest and participation over time. This paper presents our findings from a long-term, multi-site deployment of a pervasive health game, the American Horsepower Challenge (AHPC). Similar to previous studies, our findings also show reduced effectiveness of the game, but on a much larger scale. Moreover, we analyze reasons for this and report what kind of game related online and offline activities happened during the deployment. We argue that a shift of evaluation metrics and design goals is required to make real-world sustainable behavior changes. Based on empirical data, we propose three goals for AVGs—sustainability, adaptability and sociability. Behavior-changing games can learn how to achieve these goals from existing game genres, such as alternate reality games, location-based games, family games, and multiplayer online games etc.