The present paper describes the process of identifying lexical bundles, i.e., frequently recurring word sequences such as by means of and in the end of, in secondary school history and physics textbooks. In its determination of finding genuine lexical bundles, i.e. the word boundaries between lexical bundles and surrounding arbitrary words, it proposes a new approach to come to terms with the problem of extracting overlapping bundles of different lengths. The results of the structural classification indicate that history uses more NP/PP-based and less dependent-clause-based bundles than physics. The comparative analysis manages to restrict this difference to the referential function. History almost only refers to phrases, i.e. within clauses, while physics much more tends to make references across clauses. The article also includes a report on an extension of the study, ongoing work where the automatic identification of multi-word expressions in general is in focus.