Within the context of an interdisciplinary research project studying allusion in first century Latin epic, we examine some of the shortcomings of current digital methods for detecting poetic allusions through text reuse. We discuss the role that textual features above and beyond word identity often play in signalling intertextual connections to human readers, in particular, narratological structures and typical scenes. We suggest that such features can be captured by performing clustering and/or classification on word frequencies in paragraph-sized chunks of text, and propose a possible pipeline for fusing similarity measurements based on these thematic features with existing metrics produced by the intertext search tool Tesserae. While work testing the stability of the proposed features is ongoing, we report some promising initial results based on k-means clustering, and describe plans for future work.