The two fundamental challenges in Virtual Reality over the WWW are (i) to facilitate the process of creating convincing, non-trivial, highly interactive and maintainable content and (ii) to efficiently deliver this content to the client with QoS guarantees. This paper focuses on the first challenge: we have created a data model for spatio-temporal compositions that allows us to define relative placement of objects in a virtual world by means of their geometric characteristics (Joints). Thus, we model spatial relationships of objects at a higher level of abstraction than simple co-ordinate system transformations. We also define the behavior of objects in a declarative way, by means of Event-Condition-Action rules. Events and Actions can be synthesized by a set of operators for the creation of really complex behavior. Our model has been realized in prototype form in Spatio-Temporal Descriptive Language (STEDEL) that demonstrates the validity and power of our approach in terms of re-usability and authorship effort.