The web infrastructure continues to grow in both size and diversity and is quickly approaching 2 billion sites that serve an equally massive number of users. Content replication is a common technique that aims to improve reliability and performance. It consists of the use of multiple server mirrors that operate across geographically distributed regions. Several factors impair the static system optimization or even a periodic optimization. Final energy costs depend on the level of workload handled by each region given local variations in energy pricing. To accentuate the problem, energy pricing may dynamically change for some sites. Also, network state fluctuations, that are produced by congestion and failures, create time-varying performance to different users. These factors contribute to affect user experience and service costs. We propose an energy-delay aware web routing method that dynamically directs user requests to a set of replicated sites. The method relies on learning to implement a customized performance-cost load-balancing. We present experimental results from a network testbed using actual power measurements and simulated spatial variations in the energy prices according to the U.S. electricity market. The results show that the method can help to achieve the desired balance of energy cost and average response time, while reducing energy costs up to 11% without introducing a major impact to service quality.