Well that's horrifying.

For those that are wondering like me: https://blog.apnic.net/2020/08/21/chromiums-impact-on-root-d...

Essentially, it generated 3 random, single-label (so, "TLD") queries, and ran them through DNS. Whatever recurser that lands at obviously wouldn't have random junk in its cache (unlike probably any other root record) and thus had to go to the root servers for it.

They did this to try to not display an omnibox result for single words (the example in the article is "marketing") on what I'm going to just call "broken networks"; see TFA if you're curious.