I've heard that certain lisps (Common Lisp comes up when I search for reader macros) allow for all kinds of tinkering with themselves. But the ability of one to make itself not a lisp anymore, while interesting, doesn't seem to say much about the merits of sticking to s-expressions, except maybe to point out that somebody once decided not to.
Reader macros are there to program and configure the reader. The reader is responsible for reading s-expressions into internal data structures. There are basically two main uses of reader-macros: data structures and reader control.
A CL implementation will implement reading lists, symbols, numbers, arrays, strings, structures, characters, pathnames, ... via reader macros. Additionally the reader implements various forms of control operations: conditional reading, reading and evaluation, circular datastructures, quoting and comments.
This is user programmable&configurable. Most uses will be in the two above categories: data structure syntax and control. For example we could add a syntax for hash tables to s-expressions. An example for a control extension would be to add support for named readtables. For example a Common Lisp implementation could add a readtable for reading s-expressions from Scheme, which has a slightly different syntax.
Reader macros were optimized for implementing s-expressions, thus the mechanism isn't that convenient as a lexer/parser for actual programming languages. It's a a bit painful to do so, but possible.
A typical reader macro usage, beyond the usage described above, is one which implements a different token or expression syntax. For example there are reader macros which parse infix expressions. This might be useful in Lisp code where arithmetic expressions can be written in a more conventional infix syntax. The infix reader macro would convert infix expressions into prefix data.
