Here, we use geostatistical analysis to understand the spatial scales of variability in OCO-2 data and to infer the relative imprint that surface fluxes and atmospheric transport leave on the observations. We analyze nadir- and glint-mode observations from OCO-2, first focusing on tracks obtained in the vicinity of TCCON stations, subsequently expanding our analysis to broader regions in the Northern and Southern Hemispheres. Given a focus on fine-scale variations that may not be resolved at the spatial resolution of typical atmospheric transport models used for inversions, we remove the background from the OCO-2 track using a high-pass filter, and calculate semivariograms from the high frequency signal. We fit a spherical model to each semivariogram to determine the unexplained variability, which approximates the single sounding error, and the explained variability, owing to coherent differences among soundings that grows as a function of separation distance. We find that the range of coherent variations is generally between 10 and 50 km. While the range does not exhibit a strong seasonal cycle, the magnitude of the explained variations is generally largest during the growing season, when spatial gradients are maximized. For example, Northern Hemisphere explained semivariances are typically less than 1 ppm2 during the winter and fall seasons, but increase to around 5 ppm2 during the summer growing season. We compare the explained along-track semivariance with temporal variations at TCCON sites for OCO-2 tracks within 500 km of each site. We find good agreement seasonally between explained variations in OCO-2 XCO2 and mesoscale to synoptic scale variations in TCCON data, suggesting our method isolates a transport signature in the satellite observations. We corroborate this analysis with XH2O observations from both OCO-2 and TCCON stations. The larger dynamic range in XH2O permits us to further decompose the error budget on XCO2 variations, and to further explore the meteorological conditions under which fine-scale transport error is significant. We expect that these results will be useful in developing a realistic error budget for satellite observations when applied toward flux inference.