One issue of real interest in the area of web data mining is to capture users' activities during connection and extract behavior patterns that help define their preferences in order to improve the design of future pages adapting websites interfaces to individual users.This research is intended to provide, first of all, a presentation of the methodological foundations of the use of probabilistic languages to identify relevant or most visited websites.Secondly, the web sessions are represented by graphs and probabilistic context-free grammars so that the sessions that have the highest probabilities are considered the most visited and most preferred, therefore, the most important in relation to a particular topic.It aims to develop a tool for processing web sessions obtained from a log server represented by probabilistic context-free grammars.