The growing volume of heterogeneous and distributed information on the World Wide Web has made it increasingly difficult for existing tools to retrieve relevant information. To improve the performance of these tools, this paper suggests how to handle two aspects of the problem. The first aspect concerns a better representation and description of Web pages. A new concept of Web documents described by metadata is introduced; Dublin Core semantics and the XML syntax are used to represent these metadata. Ways that this concept can improve information retrieval on the Web and reduce the network load generated by robots are suggested. The second aspect is a flexible architecture based on two kinds of robots--"generalists" and "specialists" that collect and organize metadata, in order to localize the resources on the Web. The robots will contribute to the overall auto-organizing information process by exchanging their indices. Two figures present possible organizations of Web pages and interaction between specialists and generalists to answer a user need. (Author/AEF)