Here, we review the prediction accuracy for soil properties using portable X‐ray fluorescence (pXRF), mid‐infrared (MIR), and visible near‐infrared (Vis‐NIR) and the factors impacting predictions and its accuracy. In total, 305 published papers were reviewed, and most of them were from Australia, Brazil, China, and the United States. About 44% of papers focused on the prediction of soil organic carbon (SOC) using Vis‐NIR spectra. Partial least squares regression was most frequently used. Most studies sampled Alfisols, Inceptisols, and Entisols, and up to 40‐cm depth. Researcher‐based factors (type or brand of spectrometers, which differ in hardware, spectral range, resolution, and calibration protocols; preprocessing methods; prediction models; and soil analysis methods for calibration) and soil‐based factors (horizon and depth) were explored. MIR spectra had better prediction accuracy with a mean R ² over 0.8 for sand, clay, total N, total C (TC), SOC and soil inorganic carbon (SIC), and cation exchange capacity compared to Vis‐NIR and pXRF. In the past 20 years, prediction accuracy tended to increase for sand, silt, clay, SIC, soil organic matter, and EC when using MIR and Vis‐NIR spectra, and for TC and CaCO ₃ when using pXRF spectra. Preprocessing methods, spectral range, calibration, type of the prediction models (i.e., machine and deep learning), and source of soil spectra (Vis‐NIR, MIR, and pXRF), which are used to reduce noise and multicollinearity, calibrate data, and smooth spectra, all affected the prediction. In general, MIR spectra obtained the highest prediction accuracy for most soil properties. Future studies should focus on the effects of soil‐based factors (parent material, soil mineralogy, pedogenesis, soil type, and horizon/depth) on the prediction accuracy of soil physical and chemical properties.