Web prefetching is a technique that has been researched for years to reduce the latency perceived by users. For this purpose, several web prefetching architectures have been used, but no comparative study has been performed to identify the best architecture dealing with prefetching. This paper analyzes the impact of the web prefetching architecture focusing on the limits of reducing the user?s perceived latency. To this end, the factors that constrain the predictive power of each architecture are analyzed and these theoretical limits are quantified. Experimental results show that the best element of the web architecture to locate a single prediction engine is the proxy, whose implementation could reduce the perceived latency up to 67%. Schemes for collaborative predictors located at diverse elements of the web architecture are also analyzed. These predictors could dramatically reduce the perceived latency, reaching a potential limit of about 97% for a mixed proxy-server collaborative prediction engine.