Can anyone comment on how effectively privacy is preserved with this approach?
It is plausible for example that you could form a statistical cohort around people that work in an organization and determine that they have become interested in a topic that reveals internal plans.
Even as an individual, I can only see more accurate modeling of my interests as a tool to be used against me. I do not consent to sharing those correlations. They are private.
The ad agency is only able to see: their add (y), published on source z, led to x conversions, over a period of time (p).