[in response to Simon Hobson] Audio watermarks can be fairly robust
Yes, I think the servers are oversubscribed right now, and having to encode the tracks on the fly as well.
In terms of robustness, the watermark research effort I worked on spawned "PromoProtection", which EMI (and later others such as V2) used to trace ownership of pre-release promo CDs that were supplied to commercial radio stations. Each radio station needed to pay for their promo, but the problem was that once some stations got hold of it, others would quickly receive copies illegally. The fact that they had a copy was obvious as they would air it, but the real question was where did it come from?
We would make a unique serial number for each purchase that would identify the station and the track itself. Then, we had these computers tuned in to every FM and DAB radio station scanning for the presence of these watermarks. As soon as it found a station playing an illegal copy, it would also know where the copy came from and fine the offender. Within 6 months, promo piracy ended.
The watermark survived most things, including transcoding to a lossy format, from that lossy format to another, or even from a loudspeaker across a noisy pub into a PDA's microphone - you'd be surprised.
Changing the payload was also quite tricky because its locations in time and frequency were scattered. Attacks usually lead to either damaging the listening quality of the music, or simply adding additional codes to the track, all of which the decoder could spot.
I don't think Apple are using PromoProtection though, because it needs to encode in the PCM domain (even more CPU than an MDCT domain encoder).