We investigate the variable performance of a genetic algorithm (GA) on randomly generated binary constraint satisfaction problem instances which occur near the phase transition from soluble to non-soluble. We first carry out a conventional landscape analysis and observe, next to a number of common features related to the interaction structure, important differences between the instances, such as the number of solutions, the quality of the paths to the solutions, and the lengths and extent of the neutral paths for mutation. We then split the dynamics of the GA into two phases: the ascent towards the high fitness region, and from this high fitness region to a solution. To gain further information about the GA's behavior in the first phase, we construct two models based on the much simpler fully separable functions, and try to match instances which show a similar performance distribution. Although far from exact, this technique of comparing with analog search problems gives a hint about the landscape elements that are responsible for the GA's slow ascent.