I think generalized annotation has the following problem:
Draw a graph of all web pages ordered by how many times they are visited. You get a power law distribution. On the left are things like CNN's front page, on the right things like a text file containing a cheesecake recipe someone posted in 1995 that is technically still online but nobody has visited in years.
On the left side are pages that are visited too often, and public annotations inevitably descend into madness and chaos, even worse than comment sections which have the courtesy to at least be isolated on the page to a certain area. On these pages, trying to use the annotation software is worse than not using it at all.
On the right are the pages that nobody visits, so when you visit them, there is no chance of any annotations being on the page, nor any chance that if you leave any annotations anybody will ever see them, or care if they do (i.e., just giving people a directory of "rarely annotated pages" doesn't solve the problem, the problem is that nobody cares about these pages anyhow). Annotation software doesn't harm these pages technically, but the user experience is at best that the user will just forget about their annotation software, and at worst, they'll feel alone on the page, a concept not previously in their mental model and one that is not contributing to a positive assessment of the annotation software.
In between is the sweet spot, where participation is adequate that the annotation software brings some sort of value, but it isn't just overwhelming chaos.
I submit to you that viewed through the power law lens, that sweet spot is actually fairly narrow. Moreover, if you slice through a single user's browser history looking for when they hit pages in that sweet spot, it'll still be the minority of pages that they viewed, so basically by statistical necessity, the pages where the software added value must be a tiny minority of the pages you visit. The majority of pages they visit, the annotation software experience ranges from extremely net negative to at best neutral, and it's hard for the positive experiences to make up for that.
Moreover, this sweet spot is moving. As more of the general public tries out your software, the sweet spot moves to the right, which also has the effect when viewed from a single user's point of view of making the pages where it is useful become more rare. When you're just starting out and you've got a 100 users, the useful pages are things like "the front page of CNN" and "the hot wikipedia article about $CURRENT_EVENT", but as your user count increases it moves down to specific articles on CNN and only the linked content from the $CURRENT_EVENT, then only to archival content on CNN and random Wikipedia pages, and just in general into stuff that becomes increasingly difficult for it to be any significant percentage of the pages you visit.
I think this is why A: They can't get popular B: the ones that I know about that have been around for a while and are therefore presumably successful enough for someone to consider them worthwhile can only stay so provided they don't get more popular and C: there's no chance you can make money in this space because you get an anti-network effect... the more users you get, the less valuable the service becomes to every existing user!
I think this is also why it seems like such an appealing idea. You create a prototype, get a couple of friends on it, it seems like fun and to be useful. You annotate some popular pages, you have some similar interests and you annotate, I dunno, the latest game console announcement or something, and it seems like fun. The problem is, rather than this being the worst the experience will ever be as it just gets better and better as more people come online, this is the best it will ever be.
Also, I can tell you from the first couple of times around that if this did become popular, the content producers of the Internet would fight you tooth and nail. They did the first couple of times when the math sort of worked to at least get to the point that these things could get in the news. With the web so much larger and power-law-y and the anti-network effect correspondingly so much more powerful, now this sort of thing can't even be successful enough to so much as get noticed by anyone before it has already collapsed.
(I specifically said "generalized" annotation at the top, because specialized use cases can get around some of these issues. But it's going to be hard to make any money on the size of user base you can support, because while you can mitigate the anti-network effect, it's always going to be looming over you if you try to get large enough.)