The compaction quality of soil embankments is critical to the long-term performance of the pavements placed on them. In current quality assurance (QA) practice, state highway agencies (SHAs) rely on in-situ testing at a small number of point locations to decide whether to accept or reject the product, assuming that the samples taken at random locations are independent of each other. This assumption, however, is invalid because soil properties are spatially autocorrelated - the properties at nearby locations are correlated to each other. Consequently, if the sampling locations are close to each other, the effective number of samples is reduced, which in turn increases the risk of incorrect accept/reject decisions. This study addressed this spatial autocorrelation issue in soil acceptance testing. Soil data from the U.S Highway 31 project, collected by intelligent compaction (IC) in the format of compaction meter value (CMV), were used to prove the existence of spatial autocorrelation using the semivariogram and Moran's / analysis. The impact of spatial autocorrelation on soil acceptance testing was assessed by comparing the testing power under two scenarios (with and without spatial autocorrelation). The results suggest that the existence of spatial autocorrelation decreases the testing power, resulting in a greater risk to the SHA. Based on these findings, this study proposed two spatial indices to mitigate the negative impact of spatial autocorrelation by controlling the spatial pattern of random samples. A web tool was also developed as an implementation to augment the random sampling process in field QA practice by incorporating the spatial pattern of samples.