We collect Word-of-mouth (WOM) data on movies from Twitter and employ both a time-series model and a dynamic panel data model to study the influence of WOM on movie box office revenues. Compared with most previous literature that measures WOM through its volume or dispersion, we directly measure the number of recipients of each WOM message using the unique social structural information on Twitter. Thereby we offer a more direct study of WOM and provides a powerful evidence of the causal effect of WOM on product sale which is rarely dealt with in the literature. We also disentangle the different roles of pre-consumption WOM and post-consumption WOM for the first time in the literature and we find that the percentage of pre-consumption WOM has significant explanatory power. Although previous studies conclude that the valence of WOM does not have any explanatory power for movie box office revenue, our time-series based analysis suggests that valence of WOM does play an important role. Our conceptual model and empirical study shed lights on how WOM actually influences product sales and this paper also reveals the value of social networking sites like Twitter to both marketing researchers and practitioners.