Since the amount of network traffic has rapidly increased with the WWW expansion, users have experienced a long latency when retrieving Web pages. To solve the latency problem, we propose a client-side prefetching mechanism, which reflects changes on Web document structures and utilizes information from the entrance pages of frequently visited Web sites. It starts by constructing link graphs by gathering usage information of visited Web sites, and predicts the next document to be referenced based on the overall displayed documents in the Web browser. We also manage entrance pages not to be easily replaced from the cache. Our simulation results show that it has a remarkably improved performance: an increased cache hit ratio (by 48%) and a high prefetching effect (by 297%), with a slightly increased network overhead compared to similar previous schemes.