Much of the research on automated Web Service Composition (WSC) relates it to an AI planning task, where the composition is primarily done offline prior to execution. Recent research on WSC has argued convincingly for the importance of optimizing quality of service, trust, and user preferences. While some of this optimization can be done offline, many interesting and useful optimizations are data-dependent, and must be done following execution of at least some information-gathering services. In this paper, we examine this class of WSC problems, attempting to balance the trade-off between offline composition and online information gathering with a view to producing high-quality compositions efficiently and without excessive data gathering. Our investigation is performed in the context of the semantic web employing an existing preference-based Hierarchical Task Network WSC system. Our experiments illustrate the potential improvement in both the quality and speed of composition generation afforded by our approach.