I'm not sure what he's alluding to. It may be that they've done simple loudness normalisation instead of one of the more preferable, advanced techniques, like ReplayGain 1.0, R128, or BS.1770 (/ReplayGain 2.0)?
Some of these techniques appear to (optionally) take the entire album into account when conducting normalization.
Maybe that's it. Due to its nature, YouTube tends to lend itself to album fragmentation. It's not uncommon for people to have an entire album on their playlist where no two tracks were uploaded together, let alone even share the same uploader.
In this scenario, barring some sort of mitigation strategy, any seamless transitions between tracks would most likely be destroyed. Each track would have been normalized individually.
