Atomic magnetometers are sensitive tools for weak magnetic-field detection. In alkali-metal atomic magnetometers, multipass cells can extend the effective atom-light interaction length and enhance the optical readout signal. Previous studies have mainly used such structures to improve magnetometer sensitivity or to realize compact magnetic-gradient measurements. However, the influence of the multipass sampling path on the magnetic-resonance response in non-uniform magnetic fields remains insufficiently understood. In this work, we investigate the response characteristics of a Herriott multipass atomic magnetometer under magnetic-field gradients, with emphasis on transverse gradients. A spatial convolution model is established for an RF-driven dual-beam atomic magnetometer. The measured magnetic-resonance signal is described as a spatially weighted superposition of local Lorentzian responses. In this framework, the first moment of the effective sampling function determines the resonance-center shift, while its second central moment determines gradient-induced broadening. The model is used to compare the Herriott multipass mode with the single-pass mode. Experiments are performed in a magnetically shielded system under controlled transverse and axial magnetic-field gradients. The absorption line shape, resonance-center frequency, linewidth, dispersive slope, and equivalent magnetic-field noise are measured and analyzed. The results show that the single-pass mode maintains an approximately Lorentzian line shape under transverse gradients because of its localized spatial sampling. In contrast, the Herriott mode gradually evolves from a nearly Lorentzian profile into a broadened, flattened, non-Lorentzian line shape with shoulder-like structures. This behavior confirms that the Herriott response is determined by the weighted superposition of multiple local magnetic-resonance signals with different Larmor frequencies. Further analysis shows that transverse gradients significantly increase the linewidth and reduce the central dispersive slope in the Herriott mode, leading to degradation of the equivalent magnetic-field noise. Therefore, the readout sensitivity gain obtained from the multipass optical path at zero gradient is partially offset under large transverse gradients. Axial-gradient measurements indicate that the axial response is mainly determined by the effective sampling length along the cell axis rather than by the transverse ring radius of the Herriott pattern. These results reveal the trade-off between optical-path gain and gradient tolerance in Herriott multipass atomic magnetometers and provide guidance for optimizing multipass vapor-cell structures in non-uniform magnetic-field environments.