The ability to detect the identity of a sample obtained from its environment is a cornerstone of molecular ecological research. Thanks to the falling price of shotgun sequencing, genome skimming, the acquisition of short reads spread across the genome at low coverage, is emerging as an alternative to traditional barcoding. By obtaining far more data across the whole genome, skimming has the promise to increase the precision of sample identification beyond traditional barcoding while keeping the costs manageable. While methods for assembly-free sample identification based on genome skims are now available, little is known about how these methods react to the presence of DNA from organisms other than the target species. In this paper, we show that the accuracy of distances computed between a pair of genome skims based on k-mer similarity can degrade dramatically if the skims include contaminant reads; i.e., any reads originating from other organisms. We establish a theoretical model of the impact of contamination. We then suggest and evaluate a solution to the contamination problem: Query reads in a genome skim against an extensive database of possible contaminants (e.g., all microbial organisms) and filter out any read that matches. We evaluate the effectiveness of this strategy when implemented using Kraken-II, in detailed analyses. Our results show substantial improvements in accuracy as a result of filtering but also point to limitations, including a need for relatively close matches in the contaminant database.