Lexical cohesion refers to the reader-perceived unity of text achieved by the author's usage of words with related meanings (Halliday and Hasan, 1976). This article reports on an experiment with 22 readers aimed at finding lexical cohesive patterns in 10 texts. Although there was much diversity in peoples' answers, we identified a common core of the phenomenon, using statistical analysis of agreement patterns and a validation experiment. The core data may now be used as a minimal test set for models of lexical cohesion; we present an example suggesting that models based on mutually exclusive lexical chains will not suffice. In addition, we believe that procedures for revealing and analyzing sub-group patterns of agreement described here may be applied to data collected in other studies of comparable size.