It seems to me the authors make three main points
1) Pay-per-action confuses correlation (action when seeing ad) with causation (action because of seeing ad). Hence click-through-rate maximizes correlation instead of causation. Google knows you were going to buy anyway, thus serves you an ad.
2) Marketers and ad networks interests' are aligned in that they want to maximize advertisement budgets. Hence marketers will ignore randomized studies showing ads of some type are not effective.
3) Even when these studies are performed, even major sample sizes are insufficient to support any conclusion