Summary The objective of this study is to improve understanding of the geostatistics of vertical (bed-normal) permeability (kz) and its influence on reservoir performance during CO2 enhanced oil recovery (EOR) and storage. kz is scrutinized far less often than horizontal permeability (kx, ky) in most geological and reservoir modeling. However, our work indicates that it is equally important to understand kz characteristics to better evaluate their influence on CO2 EOR and storage performance prediction. We conducted this study on approximately 9,000 whole-core triaxial permeability (kx, ky, kz) measurements from 42 wells in a San Andres carbonate reservoir. We analyzed kz data, including heterogeneity, correlation, and sample sufficiency measures. We analyzed wells with the largest and smallest fractions of points with kz > kmax = max(kx, ky) to explore geological factors that coincided with large kz. We quantified these geological effects through conditional probabilities on potential permeability barriers (e.g., stylolites). Every well had at least some whole cores where kz > kmax. This is a statistically justifiable result; only where Prob(kz > kmax) is statistically different from 1/3 are core samples nonisotropic. In conventional core data interpretation, however, modelers usually assume kz is less than kmax. For the well with the smallest fraction (11%) of cores where kz >kmax, the cumulative distribution functions (CDFs) differ and coincide with the presence of stylolites. We found that kz is approximately twice as variable as kx in many wells. This makes kz more difficult to interpret because it was (and usually is) heavily undersampled. To understand the influence of kz heterogeneity on CO2 flow, we built a series of flow simulation models that captured these geostatistical characteristics of permeability, while considering kz realizations, flow regimes (e.g., buoyant flow), CO2 injection strategies, and reservoir heterogeneity. CO2 flow simulations showed that, for viscous flow, assuming variable kx similar to the reservoir along with a constant kz/kx = 0.1 yields a close (within 0.5%) cumulative oil production to the simulation case with both kx and kz as uncorrelated variables. However, for buoyant flow, oil production differs by 10% [at 2.0 hydrocarbon pore volume (HCPV) of CO2 injected] between the two cases. Such flows could occur for small CO2 injection rates and long injection times, in interwell regions, and/or with vertically permeable conduits. Our geostatistical characterization demonstrates the controls on kz in a carbonate reservoir and how to improve conventional interpretation practices. This study can help CO2 EOR and storage operators refine injection development programs, particularly for reservoirs where buoyant flow exists. More broadly, the findings potentially apply to other similar subsurface buoyancy-driven flow displacements, including hydrogen storage, geothermal production, and aquifer CO2 sequestration.