Surface nuclear magnetic resonance (NMR) is a noninvasive geophysical tool used to investigate groundwater reservoirs. The relevant physical process in surface NMR is the nuclear spin of hydrogen protons in liquid water. Standard single-pulse surface NMR experiments provide estimates of water content in the shallow subsurface. Under favorable conditions, pore-structure and even hydraulic-conductivity information can be extracted from double-pulse surface NMR data. One crucial issue in surface NMR experiments is the resonance condition: the frequency of the excitation field should closely match the Larmor frequency of the protons, which is controlled by the local magnitude of the earth's magnetic field. Although the earth's field can be measured accurately by an on-site magnetometer, several effects impede perfect matching of the frequencies. These include temporal variations of the earth's field, instrumental imperfections, and the magnetic susceptibility of the underlying rocks. We assess the impact of violating the resonance condition on surface NMR experiments. Our investigation involves numerical simulations and measurements using a sample-scale earth-field NMR device and a surface NMR acquisition system. For frequency offsets up to 5 Hz, we have find that relatively standard single-pulse surface NMR recording procedures are likely to produce reliable water-content estimates as long as the pulse moments are small to moderate or the aquifer is relatively deep. If strong pulse moments are required or shallow aquifers are probed, off-resonance conditions can lead to anomalous increases in recorded amplitudes that can be mistakenly interpreted in terms of deepwater occurrences. Double-pulse surface NMR experiments are particularly sensitive to off-resonance effects, such that the results may be highly biased even for the small-frequency offsets commonly encountered in field situations.