This is not new, nor is it restricted to millennials. There have been numerous studies over the years initially looking at the attention difference between those with computers and those without. The latest round has been looking at before smartphones and after. A couple of years back there was a press flurry of humans now less attentive than goldfish. Nor, of course, is it just about ads; it's about pretty much anything. Look! Bird!
In my little circle of contrary webdevs we call this the lowering re-attention span. Basically something needs, first, to catch one's attention and then every n-time period must re-catch one's attention aka keep them engrossed. Looking at numerous studies since the advent of the smartphone it has been decreasing at ~0.5 second per year. At some point it has to level out or attention twitch syndrome could become an extinction event.
Reminder: average means roughly as many stay longer as leave earlier, a middle not a container/spread.
This tosses a huge wrench into many folks speed considerations: my new page loads/renders blazing fast but visitors are still leaving! Are leaving even faster! It loaded, they read/scanned for n-seconds and were not grabbed aka enticed to keep reading/scanning...repeat every n-seconds.
With mobile, of course, we also have the horror of third party content, i.e. ad networks, being slow to load so that the screen can look blank - the content below initial viewport - or jerking all about as each loads in irritating - late later latest - sequence. Not a good thing regards grabbing and holding attention; more a shoving and loss of attention.
Yes, time on a given page varies: by gender (males tend to leave faster), by age (younger tend to leave faster), etc. And for added fun, what grabs/holds young/old, male/female, ethic/social group, etc. is not always the same. Which can make page, copy, media, design and layout an exercise in frustration: what was lost on the swings but made up on the roundabouts becomes what's lost on the roundabouts made up on the swings; round and round we go. And one reason why I'm increasingly delivering content on visitor context and not via a singular default page.
The web is not what it was 5-years ago which is not what it was 10 years ago which is not...
Nor what it will be in another 5.
Nor, apparently, are/will we.