The study introduces a new quantitative precipitation estimation (QPE) algorithm from Advanced Baseline Imager (ABI) observations from GOES-16 across the contiguous USA (CONUS). It is developed and comprehensively evaluated using the Ground Validation Multi-Radar/Multi-Sensor (GV-MRMS) system as a benchmark, and features Random Forest (RF) machine learning-based QPE. The key innovations of the algorithm include a comprehensive set of satellite predictors derived from five infrared ABI channels, complemented by low-level environmental conditions from the Rapid Refresh (RAP) numerical weather prediction (NWP) model, and outputs of probability of precipitation type for seamless integration of varying precipitation rates across types. A systematic analysis of the predictors is performed. The analysis reveals that satellite predictors contribute more to high-intensity precipitation estimates, whereas environmental predictors primarily condition low-intensity precipitation. Combining both categories of predictors improves scores (correlation coefficient, CC = 0.41) overall, with the greatest improvement in the warm stratiform precipitation type. Introducing precipitation type information through probabilities further improves correlation (0.46). An intercomparison of the RF model with the Self-Calibrating Multivariate Precipitation Retrieval (SCaMPR) shows that RF has better detection and quantification scores than SCaMPR (Heidke Skill Score, HSS = 0.91 vs. 0.19; CC = 0.44 vs. 0.26). Both retrievals display similar performance patterns across different regions of the CONUS. For example, skill degrades in complex terrain. Precipitation processes in complex terrain and their variability may not be well captured, especially with the degraded ABI resolution in the western CONUS. It is recommended to complement the 11.2 μm legacy channel with at least the 6.2 μm channel for global precipitation retrievals using GEO sensors. © 2022 Royal Meteorological Society.