Learning Bayesian networks from data is an NP-hard problem with important practical applications. Metaheuristic search on the space of node orderings combined with deterministic construction and scoring of a network is a well-established approach. The comparative performance of different search and score algorithms is highly problem-dependent and so it is of interest to analyze, for benchmark problems with known structures, the relationship between problem features and algorithm performance. In this paper, we investigate four combinations of search (Genetic Algorithms or Ant Colony Optimization) with scoring (K2 or Chain). We relate node juxtaposition distributions over a number of runs to the known problem structure, the algorithm performance and the detailed algorithmic processes. We observe that, for different reasons, ACO and Chain both focus the search on a narrower range of orderings. This works well when the underlying structure is compatible but poorly otherwise. We conclude by suggesting future directions for research. © 2010 Springer-Verlag.