That depends on how you evaluate the group. If you track each shot's location, then a large group evaluated by radial standard deviation is best. If you track only the group diameter as determined by the two furthest-spaced shots in your groups (group diameter) it turns out that firing 7-shot groups and averaging the diameters of several of them get's you to statistical nirvana with the fewest number of rounds fired. A paper covering this and how many groups of 7 you need to get certainty within different percentages is here. It says the average diameter of 6 groups of 7 shots (42 shots fired altogether) gets you to within 15% error expected for an infinite group count with 90% confidence.
The radial standard deviation method will do that in 15 shots with 95% confidence, so probably about 13 shots for 90% confidence, which is just one more than the number of holes you are including in your measurements of the extreme spread of six groups of 7. The radial SD method is the most ammunition-economical method, but you do need to know where each round lands to use it.