The World-Wide Web is an enormous, distributed, and heterogeneous information space. Currently, with the growth of available data, finding interesting information is difficult. Search engines like Altavista are useful, but their results are not always satisfactory. In this paper, we present a method called Knowledge Discovery on the Web for extracting connections between terms. The knowledge in these connections is used for query expansion. We present experiments performed with our system, which is based on the SMART retrieval system. We used the comparative precision method for evaluating our system against three well-known Web search engines on a collection of 60,000 Web pages. These pages are a snapshot of the IMAG domain and were captured using the CLIPS-Index spider. We show how the knowledge discovered can be useful for search engines.