After a period GP (or any other stochastic search technique) will find it difficult to improve on the best trial solution it has found so far and instead most of the trial solutions it finds will be of the same or worse performance. Selection will discard those that are worse, leaving active only those that are as good as the best-so-far. In the absence of bias, the more plentiful programs with the current level of performance are more likely to be found [Langdon and Poli1997b]. The distribution of these is similar to the distribution of trees, therefore we expect the search to evolve in the direction of the most popular tree shape.