Why would iperf-created traffic not be a good indicator of real world performance?

On pretty much all sensible networks I've tested iperf3 on, performance of actual applications sending data was close to its results.