I worked for one of the many "3D on the web" companies. Step one, of course, is to make the basics actually work reasonably effectively. Step two is to figure out what kind of content would be embraced by users.
And there was the sticking point. We tried interactive fiction, maps, greeting cards, office-style output, music players and games. Of those, high-performance video for maps was valuable (we sold off all our patents on that), and games were enthusiastically embraced by users.
Everything else was "nice", but not nice enough to be worthwhile authoring the content.
(I made a super nice interactive globe powered by an entire cheap-o pipeline from Project Gutenberg. You could point and click on a country the rotating globe, and it could give you a list of books that mentioned that country.)
There's also a question about specific technology. Of them, we explicitly decided that VRML was just plain horrible to work with: it was super hard for anyone who wasn't deep in the field to grasp and it had too big a focus on "being cool" and not enough of a focus on "being fast".