Each test ranks a list of possible commands from best (most efficient) to worst (least efficient). Lower percentages are better. You can only compare the percentages between the commands in the same test.
These tests were done in a superflat world with no entities except for the player. The only commands running are the ones from the test.