Web server logs can be used to build a variable length Markov model representing user's navigation through a web site. With the aid of such a Markov model we can attempt to predict the user's next navigation step on a trail that is being followed, using a maximum likelihood method that predicts that the highest probability link will be chosen. We investigate three different scoring metrics for evaluating this prediction method: the hit and miss score, the mean absolute error and the ignorance score. We present an extensive experimental evaluation on three data sets that are split into training and test sets. The results confirm that the quality of prediction increases with the order of the Markov model, and further increases after removing unexpected, i.e. low probability, clicks from the test set.