There is a dearth of research on cohesion in academic writing, specifically in research articles (RAs). In addition, there is little quantitative information on how cohesion is realized in different rhetorical sections of RAs. Thus, the present study investigated cohesion at sentence, paragraph, and the whole textual levels across the rhetorical sections of RAs of applied linguistics. To this end, 25 indices of local, global, and text cohesion were obtained from a corpus of 100 RAs in the field of applied linguistics divided into seven sections (abstract, introduction, literature review, method, results, discussion, and conclusion). The results of mixed-effects modeling showed that the measures of local and text cohesion were significantly affected by the rhetorical section. Moreover, random forest modeling revealed that the indices of global cohesion could predict the introduction, method, and results sections, while text cohesion was a classifier for abstract and conclusion sections. The findings are thoroughly discussed, offering insights into their theoretical and pedagogical implications.