Existing prefetching techniques rely on server-based, proxy-based, or client-based reference access information. Although Web servers may provide accurate access information, our studies show that significant communication overhead can be involved by sending unnecessary reference information to clients or/and proxy servers. Our study also shows that prediction accuracy of proxy-based prefetching can be significantly limited without input of Web servers. We propose a coordinated proxy-server prefetching technique that adaptively utilizes the reference information and coordinates prefetching activities at both proxy and web servers. In our design, the reference access information stored in proxy servers will be the main source serving data prefetching for groups of clients, each of whom shares the common surfing interests. The access information in the web server will be used to serve data prefetching only for data objects that are not qualified for proxy-based prefetching. Conducting trace-driven simulations, we show that both hit ratios and byte hit ratios contributed from coordinated proxy-server prefetching are are up to 88% higher than that from proxy-based prefetching, and they are comparable to the ratios from server-based prefetching with a difference of 5%.