Purity by (strong) default makes it easier to combine your programs. Lazyness can help, too, but I am not completely sure whether you should make lazy the default. Chris Okasaki's "Purely Functional Data Structures" argues for needing both lazy and strict evaluation to guarantee asymptotic run times.
Have a look at, say, Software Transactional Memory to see what purity can buy you. Combinatorial parsers are another interesting application in Haskell; however they should be easily implementable in Scheme, too.
Please have a look at how http://www.cs.kent.ac.uk/people/staff/smk/redblack/rb.html uses the type system to statically guarantee the balance of red-black trees.
Please pardon that I mostly talked about the `how' and not so much the `why'. I hope giving some examples might not totally be in vain.
I'd be grateful if you could expand upon that statement. I'm a Lisper, and half the time I see one of these Haskell articles, I wonder if there really is a there there. For sure, functional programming can be a whole lot better than CL or Scheme, but I'd like to know if the difference is significant.