Controlled-source electromagnetic induction is a geophysical technique used to identify anthropogenic and natural features, such as unexploded ordnance or fluid pathways, in the shallow subsurface. However, electromagnetic responses are difficult to interpret owing to the complexity of the underlying electrical conductivity structure. We show here that electromagnetic responses are fractal signals, reflecting a very rough distribution of electrical conductivity. Apparent conductivity profiles across a floodplain and a fractured sandstone aquifer both show that the fractal properties of the surface response depend on the complexity of the causative geological structure. Small-scale fluctuations in the response due to geological noise are inherently reproducible, and are not caused by random instrumental or atmospheric effects as often assumed. New approaches to modeling are required to take full advantage of the rich information content of near-surface electromagnetic data.