The recovery of metagenome-assembled genomes (MAGs) from metagenomic data has recently become a common task for microbial studies. The strengths and limitations of the underlying bioinformatics algorithms are well appreciated by now based on performance tests with mock data sets of known composition. However, these mock data sets do not capture the complexity and diversity often observed within natural populations, since their construction typically relies on only a single genome of a given organism. Further, it remains unclear if MAGs can recover population-variable genes (those shared by >10% but <90% of the members of the population) as efficiently as core genes (those shared by >90% of the members). To address these issues, we compared the gene variabilities of pathogenic Escherichia coli isolates from eight diarrheal samples, for which the isolate was the causative agent, against their corresponding MAGs recovered from the companion metagenomic data set. Our analysis revealed that MAGs with completeness estimates near 95% captured only 77% of the population core genes and 50% of the variable genes, on average. Further, about 5% of the genes of these MAGs were conservatively identified as missing in the isolate and were of different (non-Enterobacteriaceae) taxonomic origin, suggesting errors at the genome- binning step, even though contamination estimates based on commonly used pipelines were only 1.5%. Therefore, the quality of MAGs may often be worse than estimated, and we offer examples of how to recognize and improve such MAGs to sufficient quality by (for instance) employing only contigs longer than 1,000 bp for binning.