It peaked at 4.05% in March. The last 2 months it went just below 4% as the Unknown category increased. For June the reverse happened, so 4.04% seems to be the real current share of Linux on Desktop as desktop clients were read properly/werent spoofed.
Shit started hitting the fan when everyone started faking Netscape’s “Mozilla” user agent. Then “KHTML, like Gecko” and after that every fork kept the originating name in the string and extended it.
There’s a good explanation about that here: https://webaim.org/blog/user-agent-string-history/
The issue is that a lot of sites used the user-agent to determine if the browser supported particular features (e.g. show a fancy version of a site if the user is using Netscape, otherwise show a basic version for Mosaic, lynx, etc). New browsers had to pretend to be the old good browsers to get the good versions of sites
This is why getting rid of the user agent is a good thing. Sniffing the UA is a mess.