Abstract Estimates of recreational visitation are essential for public land management. Visitation is typically estimated using devices such as automated counters that require logistics and effort, particularly at remote locations or those with several access points. In this study, we investigate the utility of alternative data sources and statistical models for estimating visitation to public lands in the United States, through an analysis of data from 70 Bureau of Land Management sites. We compile 1328 site-months of visitor count data collected on-site, which are used to train and evaluate three random forest models incorporating combinations of 15 site-level characteristics and three sources of digital mobility data—mobile device locations, geolocated social media, and community science observations. Models including site characteristics perform better than models relying on mobility data alone. Cross-validation using held-out sites reveal varying prediction accuracy, suggesting that model generalizability depends on the inclusion of characteristically similar sites in the training data. These results underscore the limitations of relying solely on mobility data for visitation estimation and highlight the benefits of combining diverse data sources. Our approach provides a scalable, data-driven framework for estimating visitation where traditional monitoring is challenging or infeasible, supporting broader applications in recreation management.