In the current study, we used bootstrap analyses and the common principal component (CPC) method of Flury (1988) to estimate and compare the G-matrix of Scabiosa columbaria and S. canescens populations. We found three major patterns in the G-matrices: (i) the magnitude of the (co)variances was more variable among characters than among populations, (ii) different populations showed high (co)variance for different characters, and (iii) there was a tendency for S. canescens to have higher genetic (co)variances than S. columbaria. The hypothesis of equal G-matrices was rejected in all comparisons and there was no evidence that the matrices differed by a proportional constant in any of the analyses. The two 'species matrices' were found to be unrelated, both for raw data and data standardized over populations, and there was significant between-population variation in the G-matrix in both species. Populations of S. canescens showed conservation of structure (principal components) in their G-matrices, contrasting with the lack of common structure among the S. columbaria matrices. Given these observations and the results from previous studies, we propose that selection may be responsible for some of the variation between the G-matrices, at least in S. columbaria and at the between-species level.