Abstract Background Genome assembly is typically a two-stage process: contig assembly followed by the use of paired sequencing reads to join contigs into scaffolds. Scaffolds are usually the focus of reported assembly statistics; longer scaffolds greatly facilitate the use of genome sequences in downstream analyses, and it is appealing to present larger numbers as metrics of assembly performance. However, scaffolds are highly prone to errors, especially when generated using short reads, which can directly result in inflated assembly statistics. Results Here we provide the first independent evaluation of scaffolding tools for second-generation sequencing data. We find large variations in the quality of results depending on the tool and dataset used. Even extremely simple test cases of perfect input, constructed to elucidate the behaviour of each algorithm, produced some surprising results. We further dissect the performance of the scaffolders using real and simulated sequencing data derived from the genomes of Staphylococcus aureus , Rhodobacter sphaeroides , Plasmodium falciparum and Homo sapiens . The results from simulated data are of high quality, with several of the tools producing perfect output. However, at least 10% of joins remains unidentified when using real data. Conclusions The scaffolders vary in their usability, speed and number of correct and missed joins made between contigs. Results from real data highlight opportunities for further improvements of the tools. Overall, SGA, SOPRA and SSPACE generally outperform the other tools on our datasets. However, the quality of the results is highly dependent on the read mapper and genome complexity.