[scott00]

Trackers don't care if they're wrong some of the time. The prediction problems they're using the data to build models for are pretty noisy anyway. If using inexact identifiers improves their model, they'll get used. Many technically dynamic IPs change only rarely... I think my home Comcast IP has changed once in the last 2.5 year. So the correlation between a Comcast IP and a perfect household identifier is going to be pretty good. If you have a dataset that's got search history timestamped and labeled with IP, it's probably pretty easy to figure out the physical address that goes with the IP from map searches. Cross-reference an address to name database and now you've got a dataset with each household's (labeled by name and address, with some error) search history.