A number of factors can influence the hydraulic behaviour of faults including, the rheology and permeability of host and fault rock as well as bulk stress conditions, structural anisotropy, differences in pressure across the fault and the viscosity of the fluids interacting with fault zones. These factors ultimately control whether a fault acts as a barrier to lateral fluid flow and/or a conduit for along fault flow. This thesis examines the architecture of normal faults to improve understanding of their geometric variability and how it may impact fluid flow. I analyse six normal faults with displacements of 0.01-0.30 m from within the Late Miocene (~6.5-11 Ma) Mount Messenger Formation exposed along coastal cliffs in north Taranaki, New Zealand. For each fault the fault-zone and fault-rock thicknesses have been measured at 5 cm or 10 cm intervals and are combined with grain-size analysis of displaced beds and fault rock. In addition, the micro-structure of fault zones has been examined using thin sections and SEM images. Data have been used to help constrain the fault-zone structure and faulting processes together with their implications for fault-seal prediction. Results indicate that for individual faults both fault-zone and fault-rock thicknesses can vary by more than an order of magnitude over distances of <2 m and may be significantly larger over the entire fault surface. The wavelength of these changes may be partly controlled by bed thicknesses and associated fault segmentation. The narrowest fault zones are often observed where siltstone beds intersect the faults. The thickness of fault rock is positively related to the number of deformation bands, which form by cataclasis of host-rock sandstone. Cataclasis is regularly seen in laboratory analysis and is an important process for low permeability fault rock generation. Smear of siltstone beds is also common but not ubiquitous and in many cases, does not contribute significantly to fault-rock generation. Therefore, fault-rock thicknesses are not strongly correlated with the locations of siltstone beds. These data suggest that the utility of algorithms which primarily use shale smear in order to predict fault seal may require re-examination.