Okay, this seemed wrong. As the article said, even Win8 didn't go down in usage over time. So I went and checked the methodology for the source data.

Turns out, this number is based on social media and search engine referral data. Also turns out, they warn that while they do track Bing chat referrals when you follow through a link, they don't see chat responses where you only read the AI response but don't click through:

We have no way of measuring the number of queries performed in bing chat. However, we also don't measure the number of queries to regular search engines like bing or google either. Instead we track search engine referrals.

i.e. If you go to a search engine and do a search for anything and you click on a website result, we'll record that click as a search engine referral if that website had the statcounter code installed. It's the click to a website that we measure, not the actual search queries that were performed.

When you do a search using bing chat, and you click on one of the "learn more" websites we can track that as a search referral. So we are monitoring bing chat in the same way we measure the regular bing search engine.

From this data we can see from the statcounter network of webites, that the amount of traffic being sent to websites from bing chat is very, very small. Less than 1/100 of 1 percent.

So from our data we can say that bing chat is not currently translating into enough clicks to our network of websites to change the search share.

Of course you are less likely to click on a source website from bing chat than a regular search, as it is intended to give you the answer rather than have you go visiting websites to find the answer. So that needs to be factored in when using our stats for your analysis.

That is very interesting. That's a likely culprit for Win11 specifically to have gone down a couple of percentage points in the US and EU (the other territories seem to remain flat), but it's hard to prove.

It's also a bit concerning in terms of measuring the effects of AI search in both network traffic and in how search results are consumed. If that's the cause it does suggest that AI chat users are less likely to follow through to the source info, which seems risky, although it's also hard to prove what that does to receiving truthful info.

Lots of counterintutitive, hard to parse implications from this one data point, but I'd be surprised if it was as simple as "people have randomly decided to roll back to Win10 (and Win8, which also grows) for no reason".