It seems like what you posit is only a mystery if we take it on faith that targeting works. Once that faith is lost, you can also approach it from the angle that if showing them something we think they want is just as effective as something we know they don't want, you're actually proving that "what we think they want" isn't working either.
I mean, given "we know you've bought a vacuum cleaner" and "we think you'd like a kettle" - if ads for either are equally effective, either there's a lot more people collecting vacuums than either of us would have expected - or we're entirely wrong about the kettle. And it seems to me that our blind faith in targeted advertising leads us to wonder why people want so many vacuums.
But how do we explain how there are also advertisers like Amazon who do know that you just bought a vacuum from them and still show you more ads for it? Since Amazon is in a position to run principled A/B tests on whether showing these ads leads to sales that otherwise wouldn't have happened, and they are the kind of organization I'd expect to get this right, this part I am willing to accept without external evidence. It's probably that the likelihood of additional purchases of the same item, for yourself or others, is high enough, combined with that the cost of advertising to you is low enough.
People sending stuff they brought back and buying something else exist, and also people buying for their friends. But if you had an actually good prediction of their behavior, you wouldn't need to proxy it by "brought a vacuum recently".
Yet, companies insist that failure is a feature, and that their local maximum is the best possible world. And will keep harming advertisers, viewers and society to keep their position at that peak.
Removing PII will hopefully hurt the greedy ad serving companies most. Though I am scared they will find workarounds.
Almost everything we currently decide when you visit a page, the only real concession is that it's scoped to the current page, rather than every page that we know you've visited.
I'm honestly not convinced our current idea of targeted ads actually works. Continually showing me links to a product I bought someone for Christmas a month ago isn't remotely clever, despite being precisely targeted. "People shopping for guitars on Saturdays are more likely to convert to sales, than people shopping while they should be working" does not require storing PII, and is (hypothetically) more useful than "If you bought a guitar last week, surely you're more likely to buy a guitar this week".