Concreteness is central to theories of learning in psychology and organizational behavior. However, the literature provides many competing measures of concreteness in natural language. Indeed, researcher degrees of freedom are often large in text analysis. Here, we use concreteness as an example case for how language measures can be systematically evaluated across many studies. We compare many existing measures across datasets from several domains, including written advice, and plan-making (total N = 9,780). We find that many previous measures have surprisingly little measurement validity in our domains of interest. We also show that domain-specific machine learning models consistently outperform domain-general measures. Text analysis is increasingly common, and our work demonstrates how reproducibility and open data can improve measurement validity for high-dimensional data. We conclude with robust guidelines for measuring concreteness, along with a corresponding R package, doc2concrete, as an open-source toolkit for future research.