But that targeted ad, which as you say may cost 35 times more per impression is still much CHEAPER than the untargeted mass audience ad, right?

It depends on the circumstance. If your targeting is more than 35 times as accurate than the 35x increase in cost per impression, it's worth it. If your targeting is less than 35 times as accurate, it's worse. For example, maybe you're targeting those teenagers because you're convinced that they're the only ones for your pimple product, but have forgotten about expectant mothers with acne. Or you're getting too many teenagers with clear skin and a different targeting method would have worked better. (a fondness for Magic the Gathering, perhaps?)

Segment targeting has fallen out of fashion somewhat, and adaptive closed-loop measured targeting have pulled to the foreground. Targeting more of what has actually been working so far, letting the algorithm figure out what the "segment" is.

I worked on this stuff in the very early days, so everything was new. Among other things we were looking at pages visited, and their predictive power of the segment you're in, since ad buyers were targeting only segments at the time. The most predictive page for being female, for example, was something like brides.com/planning/hairdos. The most predictive page for being male was something to do with northeast US real time doppler weather maps. It's almost sad to think that stereotypes can be so easily confirmed.

Jim