The internal truth is that these policies are policies, but they're maintained by human beings. Human error can occur resulting in the mixing of data. In general, that would be an error and corrected as soon as it's discovered. It's also hard for those errors to occur (data is siloed, and the process of correlating between silos is designed so that someone has to eventually submit an auditable document that outlines what data they want to correlate, why they need to, and why that's the right solution in the space of alternatives).

Regarding the spooky coincidence effect: the truth is that people are a lot more predictable than they assume they are. More often than not, when you're doing a thing related to X and then 10 minutes later you see an X ad, it's because you're actually in the bucket of people who like X and the buckets are highly correlative. You might be in that bucket because you like Y and 95% of people who like Y like X, and a search for Y, followed by doing something related to X, triggered the X ad because of the Y search. There's also huge confirmation bias on those coincidence ad views; everyone has a story of one, but I've yet to meet anyone who said "I sat down and recorded every ad I was shown for a week. 90% of them correlated with an activity I was doing in an unrelated Google product between 5 and 10 minutes before the ad was shown."