A big draw is the community. Lisp has had a draw for the most ambitious and bright developers since the language was invented. Wherever researchers are trying to solve problems that have never been solved you are likely to find Lisp, as in artificial intelligence (AI) research, computer vision, planning, knowlege representation, and complex heuristic optimization. The language lends itself to solving problems both from the bottom up and the top down at the same time, which seems to help in confronting the hairiest challenges.
The exensible syntax via macros means that there is seldom a need to extend the language definition. Much of what would require a language extension in a more restricted language is just a macro away with Lisp. So Lisp programmers are free to make use of newly invented language concepts without a new language standard and without necessarily a real speed penalty. On a basic level, reams of boilerplate code are made unnecessary by small extensions. Whole new ideas in control flow, like Prolog style unification, are implemented efficiently and compactly as extensions.
The OOP system, CLOS, is in a class of its own in terms of flexibility. It is very difficult to go back to rudimentary C++/Java/C# OOP after getting a taste. GoF 5 design patterns become unnecessary as they can be expressed simply and directly.
The language has had no single corporate owner and no single definitive implementation, though it does have an ANSI standard with many conforming implementations. Major new implementations come along every decade and the old ones are still quite active. Experts can plan to use their specialized knowledge for a long time to come. This does cause some anarchistic friction and community fragmentation, but it also means the carpet cannot be pulled out and the language can't become moribund for corporate or project political reasons. There are always multiple commercial and open source implementations being worked on. The more performance focused ones regularly benchmark within a 2x factor of the very fastest, heavily funded imperative language implementations.
The Achilles heel of early Lisp commercialization was memory footprint to accomodate both the type safety features of the language and the advanced software development environments they included, with incredible features like full online documentation including graphics. A 64 MB Symbolics Lisp Machine was not viable cost-wise against an 8 MB Sun workstation. Today, RAM prices have collapsed and there is tremendous interest in the Lisp languages especially considering that the mainstream Java, C#, PHP languages today have advanced only minimally over those of 30 years ago.
There are modern languages now in competition with Lisp for mindshare with intelligent developers: Python, Lua, Erlang, Haskell, and OCaml. But none offers the same mix of maturity, adaptibility, multiple standards-compliant implementations and speed.