Semantic representations are the basis of many cognitive functions, including language production, reasoning, or creativity. To account for this fact, cognitive models increasingly draw on large-scale language embeddings derived from text or large-scale databases of semantic behavior. One shortcoming of these approaches is that they implicitly assume that everyone possesses identical semantic representations, although fundamental theories of human learning and a growing body of empirical work tell us that semantic representations must differ between individuals. Toward overcoming this limitation, this project uses simulation analysis to identify behavioral paradigms best suited to accurately measure individual-level semantic representations. Our simulation creates individual representations by permuting a popular pre-trained language embedding (fastText) and it generates responses in three semantic paradigms–free associations, relatedness judgments, and spatial arrangement—under a variety of different settings (e.g., number of cues). Based on the generated responses, we then simulate how the paradigms recover microscopic (word centrality, word similarity) and macroscopic properties (connectivity, average clustering) of the individual ground-truth representations. Our simulation shows that study designs used in past work can accurately measure some properties of individual-level representations, but not others. Importantly, we find that to measure all aspects of individual level representations reliably, studies that are extremely laborious and costly are needed. We close by discussing implications for future efforts to accurately measure and account for individual-level semantic representations in models of cognition.