Steganography has its origins in the classical "null cipher," in which a message is hidden inside an apparently innocuous, but carefully constructed, message. (For example, the plaintext could be the third letter of each word in what looks like a normal message.)
More modern versions take advantage of the fact that images (some other file types could also be used) can be modified significantly without any apparent change to the viewer. So the bits of the plaintext could be placed in, say, the low bit of each red, green, or blue color intensity. This almost certainly won't change the visual appearance of the image, but the plaintext can be easily extracted by somebody who knows that it's there.
To use this in practice, one would presumably PGP-encrypt the plaintext first and then hide the PGP-encrypted message in an image using these steganographic techniques. What the steganography does is to keep people from even realizing that there is a secret message there, generally preventing attempts at cryptanalysis or even traffic analysis.
Recent programs are said to use techniques devised so that common statistical tests on the modified image won't look any different from the results of those tests on a normal, unmodified image. This would make it especially difficult to discern which images on the web or in Usenet might have steganographic content. (I haven't looked at the techniques used in programs like outguess, but it's all open-source, so you can see for yourself what they're doing if you're of a mind to. Check out the outguess web site, for instance, for more information.)