PREMISE: Robust standards to evaluate quality and completeness are lacking in eukaryotic structural genome annotation, as genome annotation software is developed using model organisms and typically lacks benchmarking to comprehensively evaluate the quality and accuracy of the final predictions. The annotation of plant genomes is particularly challenging due to their large sizes, abundant transposable elements, and variable ploidies. This study investigates the impact of genome quality, complexity, sequence read input, and method on protein-coding gene predictions. METHODS: The impact of repeat masking, long-read and short-read inputs, and de novo and genome-guided protein evidence was examined in the context of the popular BRAKER and MAKER workflows for five plant genomes. The annotations were benchmarked for structural traits and sequence similarity. RESULTS: Benchmarks that reflect gene structures, reciprocal similarity search alignments, and mono-exonic/multi-exonic gene counts provide a more complete view of annotation accuracy. Transcripts derived from RNA-read alignments alone are not sufficient for genome annotation. Gene prediction workflows that combine evidence-based and ab initio approaches are recommended, and a combination of short and long reads can improve genome annotation. Adding protein evidence from de novo assemblies, genome-guided transcriptome assemblies, or full-length proteins from OrthoDB generates more putative false positives as implemented in the current workflows. Post-processing with functional and structural filters is highly recommended. DISCUSSION: While the annotation of non-model plant genomes remains complex, this study provides recommendations for inputs and methodological approaches. We discuss a set of best practices to generate an optimal plant genome annotation and present a more robust set of metrics to evaluate the resulting predictions.