<p>This article reviews Loet Leydesdorff’s contributions to science mapping. It explains how over the years, his mapping techniques evolved from journal mapping to global maps of science and finally towards interactive interfaces portraying multiple classifications and ontologies. It then critically reviews the challenges faced by current approaches to science mapping, which implicitly assume a ‘natural’ epistemic structure, with examples from two recent case studies. We observe that bottom-up algorithmic approaches, either based on citation or semantic approaches, lack conceptual consistency regarding the type of categories used: in a same classification a category captures methods, another one has materials, a third one contains empirical objects and a fourth is focused on theories, rather than having a single logic. I argue that science mapping would produce more useful representations by using ontologies based on a single logic that aligns with the particular conceptual needs of the analysis. Novel classification methods based on machine learning and language models hold promise to produce these tailored, question-driven ontologies.</p>