Many proposed cohesion metrics focus on the number and types of explicit cohesive ties detected within a text without also considering differences in the ease or difficulty of required referential and connective inferences. A new cohesion measure structured to address this limitation is proposed. Empirical analyses confirm that this new measure performs similarly to existing measures when applied to the simpler problem of detecting lower and higher cohesion versions of the same text, yet is significantly more effective than existing measures when applied to the more complex problem of distinguishing different texts rated by human experts as requiring lower or higher levels of reading comprehension skill. Implications of these findings relative to the goal of helping teachers and other educators select texts for use in instruction and assessment are discussed.