"Unpopular" needs to be further qualified. Rather than ask, "Why is Lisp unpopular?", ask, "Why is Lisp unpopular for X?". Lisp is actually reasonably popular for some things, and it's those things to which existing dialects are best adapted. Being well adapted to Y often means not being especially well adapted to X, even if it's for subtle reasons.

It's like asking, "Why are Arctic foxes (Alopex lagopus) not very popular on the African savanna?" Because they didn't evolve in response to the demands of that environment. The bat-eared fox, Otocyon megalotis, did.

The differences between Alopex and Otocyon aren't so massive that they seem like totally different kinds of animals. But it would be tricky to predict exactly what they should be, and develop the right kind of fox in a lab. Some adaptations are obvious (thick fur), other less so (specialized circulatory system). I suspect the same is true for programming languages. It's tricky to determine exactly what makes Scheme better suited to the classroom than to developing general applications. Perhaps the best way to solve the problem is the way PG is doing it with Arc, which is to take a very young, malleable dialect of Lisp and plunk it down in the environment to which you want it to become well adapted, and then continue development in response to the demands you encounter.