Is this even possible? Won't their ML ad models still have seen data from when they DID track people across sites? And even if they stop using those models, won't the structure and tuning of those models have been derived from cross-site-tracked data? This sounds a lot like a computer vision company saying "we have stopped using imagenet in our model training!". Did you really re-train from scratch? Ok, even if you did, now did you also happen to re-derive your network architecture from your own data?