Style: I'm embarrassed to admit that I laughed out loud at some of the jokes embedded in the code examples. The writing is clear and entertaining (why are Lisp books better in this respect than those of of other languages?).
Audience: This should probably not be the first programming book that you read, but it could easily be the second. Graham describes Common Lisp in detail, but assumes no prior knowledge of the language. This is a good book for people learning Lisp independently, for any application. Intermediate-level programmers will benefit from seeing Graham's Lisp style, which emphasizes building utilities to create a 'language' suitable for your problem.
Organization: The strongest point. Examples are keyed in well with the text: binary search trees in the data structures chapter, string substitution in the I/O chapter, ray tracing in the numbers chapter, etc. Okay, sure, there's nothing fancy there; obviously writers choose relevant examples. The impressive thing is how the examples are high-quality Lisp programs of the sort that might actually be used, even the ones from the early chapters (before the entire language is available). This is not the most common pedagogical approach, but it works here.
Possible shortcomings: There is nothing wrong with the problems per se, but most of them can be solved with very short programs. There are some great large-scale programs towards the end: an roll-your-own object system, an HTML generator, Lisp-in-Lisp; but on the other hand, you're on your own when the time comes to think of projects to try yourself.
As far as the reference section goes, it's okay, but why not just use the HyperSpec?