That first line is the reason something like this is doomed to fail. The second is that the people who would actually use annotations for something constructive are greatly outnumbered by people who will use it to post the same comment spam we see in regular comment sections. Without that being solved this is just another spam delivery vector.
I'd imagine you could start by whitelisting domains. This wouldn't help commercial sites but if you "followed" a domain it would be no worse than Twitter notifications.
If you have to start whitelisting participants, you have a very hard time creating a critical mass across the internet.
Annotation in a weird way is even harder than a "boil the ocean" problem; you have to have the exact right number of participants, across all the enormous complexity of the Internet, or at least, a very huge chunk of it. If you have nobody on some page, the annotation system is useless, and if you have thousands of people pounding away at the same page, it's pretty useless too. You have to hit this window of reliably having just the right number of users on a page, on an Internet that intrinsically tends towards power law page viewing, somehow across the entire Internet. After ~18 years of watching the space, I simply believe the entire vision is unsolvably flawed, especially in the presence of successful link aggregation and commentary sites.
It seems like something that should be constrained to social network circles. It might be useful to post graffiti on any site on the Internet - within my group of friends.
Unless your friends are very, very, very similarly minded, that falls under the "everywhere you go, nobody is there" class, and you quickly stop using it in favor of some sort of centralized comment stream like a chat channel or Facebook posts.
The power-law distribution of web pages wreaks havoc on almost any attempt to distribute people well here. Your solution basically would make it so the annotation works OK on only those pages that are on the high end of the local power law distribution, but that is probably fewer pages than you might think as you'd all have different long tails.
I don't think it's unsolvably flawed. It's just hard as hell. If you take a structured data approach to annotation and commenting, you can use analytics to weed the naturally-occurring power law distribution of comments down to just the best ideas on any particular sentence, for example.
Then you fail to activate one of the most important aspects of community building when people post comments, which is people feeling like they're participating and making a difference because other people respond to them. Take that engagement away and you're going to have a hard time building a user base.
The problem isn't that any given aspect of the annotation problem is unsolvable. The problem is that you can't solve them all at once to build a successful system.
Recall that I said I've been "watching this space" for 18 years now. That's no exaggeration; I was the first to hack Third Voice back in the day. It's not like people haven't been trying. Speculating about how it might work someday is really kind 10 or so years too late. In 1999, theorizing about how cool annotations might be was adequate. In 2017, a theory about how cool annotations might be needs to not just explain how they might be cool, but why all previous efforts have failed, and why yours might succeed. That's a much higher bar to leap now.
And one obvious solution is to trim away bits of the problem, but you end up in the local maximum of "link aggregation site with comments" pretty darned fast.
