Compressed textures always(*) have fixed compression ratios. Basically all of the formats I've seen take a tile of texels (such as 4x4) and store just enough information to regenerate the tile, or something that resembles that tile closely enough. Comparing different compression methods you have a tradeoff between blurriness and noisiness, but the basic principles (patented, of course) seem to be the same everywhere.
Paletted textures can also be seen as a primitive compression method.
The positive side of these kinds of compression is that it's (relatively) easy to build hardware that can perform random access reads from it. If you'd store your textures as PNGs or JPGs in memory, you'd have to decompress, on average, 50% of the image on every texture fetch. That's just not feasible =)
Due to these limitations, the "compressed" textures can be compressed further with traditional compression methods, like ZLIB
(*) today, at least. Who knows, maybe someone will think of something new.