During version 4, the artifact level up menu was optimized. We now get to see recommended artifact sets as well (and pick rate): showing which 2p/4p combos are used as well as the main stats for the sands/goblet/circlet. The way they sampled is from the most 'recent' players. There may be more to the sampling statistics too. For example, new players who just got a character may not have the proper artifacts for them. So some algorithm may be needed to filter out and only sample players that have sufficiently well-built characters. How this is defined may take some thought.
One way to do this is to select a sample of players that used the character to successfully clear floor 12 of abyss (say in the past 30 days or maybe even rolling 3-6 month window, depending on how much data there is). This can immediately filter out the players that at least put some time into building the character. This sample is probably large enough? How many Genshin accounts are there worldwide? If there are 10k accounts that cleared floor 12 of abyss, then the sample is probably large enough. One can then construct appropriate confidence intervals or use whatever statistical methods. And if one wants to be more selective with the sample, one can select those that got 9 stars in floor 12 or all 36 stars in the spiral abyss.
Perhaps one may also account for how different characters may be used more or less to clear floor 12. For example, Bennett may be used more for floor 12 clears compared to some other characters. Moreover, it may be that hyperbloom teams are used more often than some other teams to clear floor 12. What happens is that when showing the statistics to users, the underlying population size (denominator) may not be the same for each character, or if it is then some way or standardization method to make the denominator the same for all characters is needed (for example take 10k accounts for each character, or take the minimum of all characters).
Moreover, the underlying population may not be the same for each character. The artifact statistics for each character may be based on a different sample each time (rather than the same sample) as one cannot expect the same group of people to have the same exact characters. Some people will have certain characters, while others will have different ones. And accounts that have all Genshin characters are in the minority.
Of course, in the end, the sample may just be a cherry-picked group of people. We may not know unless we look into the source code, ask Genshin devs themselves, or spy on them.
The artifact set stats are more useful for characters that can be built multiple ways. For characters like Venti, it is expected that most players will build him with the vv set. But for other characters like Ganyu, that can be built melt (4p WT) or freeze (4p Blizzard), it may be interesting to see the artifact set distribution rates to see what 'others' are doing.
And thus, character builds can become social experiments. Does seeing how other people build a character influence the decision on how a new player builds a character? For example, if a newbie sees that say 60% (just making up the number for sake or argument) of people build Ganyu on 4p WT, would that player be more likely to build Ganyu on WT as opposed to Blizzard? And if the trend continues and more and more new players build Ganyu on WT, would we eventually reach the stationary distribution / stable equilibrium point where almost all players build Ganyu on WT?
Or could it be that Genshin leadership is guiding the playerbase with an 'iron hand', selecting the samples to encourage more players to build Ganyu in a certain way as opposed to another?
But there are other dynamic social factors that could affect pick rate (assuming it is not cherry picked from the playerbase and a bit more natural). Genshin content creators or site owners can publish videos or guides recommending certain sets to build for characters. This may or may not have an influence on what players ultimately decide to build, or it may have varying degrees of influence.
But also the abyss (one of the few forms of end game content in Genshin) changes cyclically. Certain abyss configurations may favor certain builds over others. For example, a floor 12 with bosses and enemies that cannot be frozen may favor a melt build for Ganyu more, which could then affect pick rates for artifact sets. By contrast, a floor 12 with mobs and a blessing that shreds 60% def for frozen enemies may favor a freeze build for Ganyu more.