Two approaches are commonly used when calculating the power potential of new geothermal opportunities: the US Geological Survey heat-in-place method, which requires inputs that may be challenging to determine for new prospects, and the power density (PD) method, which yields analogue-based estimates from reservoir temperature, production area, and a tectonic setting. This study evaluates how the PD method performs as an estimator given its minimalistic approach. Statistical tests show that the tectonic classification scheme for PD does not produce fully distinct subsets within the global analogue data set. In addition, the magnitude of data misfit for 100 geothermal fields suggests that the published PD curves may not be well-suited for resource prediction. Complementary geothermal field information is available in open-source geothermal databases and published literature. When modeled using transparent and traceable machine learning methods, predictions from a more feature-rich data set outperform, based on coefficient of determination and other metrics, the two-curve PD model for field-aggregated data. Even better predictive ability is demonstrated using data at the power plant scale. Drilling length, a proxy for reservoir depth, plays a crucial role in these estimators, consistent with the global average geotherm describing an increase in temperature with depth. Rift zones are also identified as a geothermally distinct tectonic setting, which contrasts with the subset of “hot arcs” highlighted by the PD method. These findings suggest a need to reevaluate the use of map-based methods like PD for power estimation, particularly when the complexities of the subsurface take three dimensions to fully describe.