>It is like the difference between a salesman asking a bartender what type of drinks you prefer and them having a transcript of everything you ever said in the bar.

What? That's a difference in scale, not in kind.

Correct. The better analogy is the salesman asking the bartender that when someone orders X drink, give them Y drink also for free. The salesman never actually gets to find out what customers got Y, though they can learn how many Y were given away and thereby get some secondary signal on how popular X is (and, of course, if individual bar patrons walk directly out of the bar and say "Hey, can I buy more Y?" the salesman can make a pretty strong guess that the patron in question likes X).

Yes. And as we know, Prolog is turing complete. Therefore†, anything which the salesman could do if he could hear ever bar patron's conversation, he can do by a clever set of ad deployments (though it may be orders of magnitude less efficient).

----

† Yes, I'm stretching the "proof" beyond its breaking point here.

It's too simple to say they don't learn anything. If you actually click on the ad, advertisers sometimes gain some information (based on IP address, third-party cookies, and maybe browser fingerprinting) Sometimes they might get some information just by running the ad via analytics.

It's a bit murky whether the amounts to being able to identify you. It's not clear how much advertisers care about identifying people who don't buy or how effective advertising actually is. And this data goes stale pretty quickly anyway.

So it comes down to, "well, it depends." None of this is comforting to people who don't want to be tracked. On the other hand I haven't heard anything credible about people being directly harmed by it? It seems more about the principle of the thing?

You're absolutely correct. That's what I was trying to capture in the parenthetical; clicking on an ad is like wandering directly out of the bar and straight into the salesman's shop.

One tells you something relevant to what sort of domain - what to get them. The bartender doesn't tell them about your nasty divorce, trouble at work, etc. Likewise Google doesn't just hand out your embarrassing search history to the highest bidder - because both would be both wrong and absolutely terrible for business.

Large differences of scale are difference of kind. This holds true for information (it's crucial that e.g. property records be available at your local planner's office; it's not clear things are similar when talking about large centralized databases that are searched from anywhere). It also holds true for systems, especially biological systems. (There's a great book called "Scale" that talks about this.)