Errors in protein synthesis can lead to non-genetic phenotypic mutations, which contribute to generating a wide range of protein diversity. There are currently no methods to measure proteome-wide amino acid misincorporations in a high-throughput fashion, limiting their detection to specific sites and few codon-anticodon pairs. Therefore, it has been technically challenging to estimate the evolutionary impact of translation errors. Here, we developed a computational pipeline, integrated with a novel mechanistic model of translation errors, which can detect translation errors across organisms and conditions. We revealed hundreds of thousands of amino acid misincorporations and a rugged error landscape in datasets of E. coli and S. cerevisiae. We provide proteome-wide evidence of how codon choice can locally reduce translation errors. Our analysis indicates that the translation machinery prevents strongly deleterious misincorporations while allowing for advantageous ones, and the presence of missing tRNAs would increase codon-anticodon cross-reactivity and mis-incorporation error rates.