> On the web, the user is rarely a monolith.

Usability folks have understood this for decades. Alan Cooper was writing about defining multiple separate personas [1] to represent different cohorts of your userbase in the 90s.

> what I'm getting at is that it is less empirical than it might seem at first pass.

I would argue that it is still exactly as empirical. You just have to be careful how you aggregate your data and don't try to reduce things to too few clusters. Otherwise you end up making the classic mistake of offering a single T-shirt size at your conference that mostly only fits men because they are the majority of attendees.

> There is still an art to user centered design,

Agreed. No amount of analysis will do your synthesis for you. You still have to make.

[1]: https://en.wikipedia.org/wiki/Persona_(user_experience)