This paper presents an innovative application of concordancing to the study of comprehension based on the hypothesis that the presence of noise in raw concordance data brings to light the existence of strings liable to cause unsuspected comprehension problems to language learners. The paper explains how the hypothesis was arrived at before proceeding to describe and discuss an experiment specially designed so as to test its validity. In the course of this experiment eight post A-level students and two native speakers/teachers of French were observed processing the raw output of a concordance of the string bien que. Details of the concordance data are given, broken down in broad categories (concessive conjunction bien que, resultative conjunction si bien que, adverb + completive, etc.). The findings are reported in three stages corresponding to those of the experiment: a pilot study involving two of the students, a repeat of the experiment with the two native speakers, and a second learner study with a larger, more representative sample of six students. Between them the results are found to provide empirical evidence in support of the hypothesis. The main outcome is that contrary to the native speakers, none of the post A-level students was able to interpret all the citations correctly. Other significant results are that difficulties were spread unevenly across citations and were more common among those with lower grades. The implications for language pedagogy are then examined. The paper concludes that concordances can be helpful in predicting potential pitfalls and ends with some suggestions for future research.