Ancestral recombination graphs (ARGs) represent the history of portions of a genome with recombination. Attempts to infer ARGs have been hampered by the lack of an ARG comparison metric which could be used to measure how well inference succeeded. We propose a simple ARG comparison framework based on averaging standard tree comparison measures across sites. Using simulated data we show that this framework, instantiated with an appropriate tree comparison measure, can distinguish better from worse inferences of an ARG.