One thing I've noticed: very few ads are topic focused. Take an ad for a car, or for life insurance. This ad isn't aimed at someone who likes cars, or insurance

One thing I've noticed: very few ads are topic focused. Take an ad for a car, or for life insurance. This ad isn't aimed at someone who likes cars, or insurance. It's aimed at an average person, within a certain income bracket and stage of life. How would you know, based on browsing history, whether a person should be displayed such an ad? If they've visited websites for insurance, or cars, then they're a shoo-in, granted. Most people haven't. So the approach will be to identify people of a certain age and income, and figure out what websites they visit. Next, use tracking cookies to identify people who visit the same set of websites. Display the ad to all such people. The original FLoC fit into this model well. Whereas it might have picked out topics that a person likes, it would also likely have picked up other bits of demographic information like age, sex, wealth, which would be more useful for targeting. In moving to an explicitly topic focused model, google throws out this capability. Presumably, people from all age groups are interested in rock music, say. But the way that different people engage with that interest will be different, and therefore offer opportunities to identify more specific groups. A purely topic based model commingles these groups, and therefore makes targeting harder. To the degree it works, it will be against the intentions of the system, since explicit interests will be used as a proxy for demographic information, rather than used directly. As such, I doubt this model will leave anyone very happy. Advertisers will dislike that it offers poorer tracking, and users will dislike that it leaks demographic information. I for one am satisfied with an approach that screws over advertisers. However, I suspect google won't be. What I see as their only escape is to create sandboxed environments in the browser from which to run advertising targeting code. A javascript worker of some sort would be restricted from making web requests, and would only be able to act on information fed to it. This information would come from webpages the user visited, and would include information about the page, as well as their actions there. The worker would use this information to decide what ads to target at the user. Downloading a particular ad without leaking information remains a problem. You could download every possible ad, and have the worker select one, but this is very wasteful of bandwidth. "Private information retrieval" (PIR) schemes can reduce the bandwidth wastage substantially, without leaking any information to the server. These schemes have received a great deal of attention from cryptographers, but as far as I can tell have been overlooked by programmers. Likely this is because there are few instances where a project cares about privacy, but the project owner is untrusted to protect privacy. Thus, it seems like privacy protecting targeted advertising would be an ideal testbed for PIR. It would be very interesting if google went this route.