This example is brought up all the time - "I just bought something which I buy every few years at most, and now I'm getting ads for it after I bought it".

Obviously it's usually completely ineffective. BUT I think it might make a purchase correlation score go up. Imagine that no human working for the company buying the ads is fully understanding what the correlation score means, and they are effectively using or subscribing to a machine learning model that is trying to increase the correlation score. And they're seeing the score go up.

I'm no expert in the ad targeting ecosystem, and I'm sure I'll be told that's not what it does. But I'm more familiar with systems software, databases, distributed services and such, and I've seen many things that didn't really do what they were supposed to do. It just takes a lot of smarts and a lot of focus, such as Aphyr does with distributed databases, to show that they don't actually do what they claim, due to some mundane bugs or subtle design flaws.

I think there's a good chance that the complex ad targeting ecosystem is just fooling advertisers into thinking it works, without actually working.