Semi-empirical models based on in situ geotechnical tests have been the standard-of-practice for predicting soil liquefaction since 1971. More recently, prediction models based on free, readily available data were proposed. These “geospatial” models rely on satellite remote-sensing to infer subsurface traits without in situ tests. Using 15,223 liquefaction case-histories from 24 earthquakes, this study assesses the performance of 23 models based on geotechnical or geospatial data using standardized metrics. Uncertainty due to finite sampling of case-histories is accounted for and used to establish statistical significance. Geotechnical predictions are significantly more efficient on a global scale, yet successive models proposed over the last 20 years show little or no demonstrable improvement. In addition, geospatial models perform equally well for large subsets of the data—a provocative finding given the relative time- and cost-requirements underlying these predictions. Through this performance comparison, lessons for improving each class of model are elucidated in detail.