I'm very skeptical of this methodology, since it's completely correlational. Here's how I would run this study as a publisher if I used an ad network that lets

I'm very skeptical of this methodology, since it's completely correlational. Here's how I would run this study as a publisher if I used an ad network that lets the publisher turn off personalization on a per-pageview level (like Google's ads, which I work on: https://support.google.com/admanager/answer/7678538?hl=en https://support.google.com/admanager/answer/7678538?hl=en). I would run an A/B test of personalized ads vs non-personalized. This is simple, and very robust since it's controlled. My expectation is that if a big US news publisher tried this they would see much more than a 4% difference, but that's speculation on my part.