Abstract Evapotranspiration (ET) and photosynthesis are key processes in ecosystem functioning on which soil moisture (SM) has an important influence. Eddy covariance measurements and machine learning (ML) increasingly enable flux prediction in ungauged regions. With numerous SM estimation methods available, each representing different spatiotemporal scales, understanding how data choice influences ML predictions is important. Our study examines how different SM data affect ML predictions of ET and photosynthesis. At semi‐arid to arid sites, we found that in situ near‐surface SM enhances ML predictions of ET. For photosynthesis, SM memory, indicative of deeper SM control, shows the highest predictive power, improving predictions by up to 30% at the driest sites. These contrasting responses reveal that ET and Gross Primary Productivity (GPP) are likely governed by distinct SM mechanisms: spatial scale matching for ET and temporal depth for GPP. Our study demonstrates that process‐guided feature engineering can improve ML predictions where root‐zone SM observations are often unavailable.