> Using this system, we ran an experiment where we turned off the availability of personalized data for a small fraction of randomly selected users through the

> Using this system, we ran an experiment where we turned off the availability of personalized data for a small fraction of randomly selected users through the programmatic arm of Google Ad Manager’s serving system, thus effectively disabling access to cookies This means the cookies for the auctions of a control group of users were not exposed. The experiment is still wrong, the bidders will still react in real time discarding the traffic without identifiers; specially since it comes from ADX where the identifier is the key to evaluating the model. Not taking any other ad exchange into account is a huge flaw if you want to extrapolate the conclusions into conclusions for the whole industry. Other exchanges provide different identifying properties such as the 4th octet of the ip_address with each auction. Google ADX doesn't send this bit of information, they obfuscate the ip_address always. It's obvious that if you remove cookies from ADX, there is no reason to use ADX since the cookie is the proxy for your model on that exchange, but on the other ad exchanges you could survive just fine. There is no way I read this experiment in which it makes sense as a conclusion for the whole ad industry.