Google's Privacy Sandbox

Google’s Privacy Sandbox is a fascinating solution to the depreciation of third-party cookies. User profiling is done directly in the browser, and only aggregated data leave the browser, ensuring privacy by avoiding 1:1 identification.

Today, I gave a presentation at PPC Camp by uLab about how this technology actually works and its implications for us, advertisers and what is my prediction for the future.

Enjoy the summary:

Targeting Options

The Privacy Sandbox works only on dektop Chrome browsers and Chrome Android apps. Not in the Chrome iOS app.

You can basically target by context or by interest groups.

Contextual Targeting

This uses only the publisher’s first-party data, such as the placement or ad slot. Publishers and ad networks cannot learn about their visitors' ad interests since this information is stored in the browser. However, they can use their first-party cookies to analyze behavior on their domain and perform their own small-scale profiling.

Publishers will (already are) pushing users to accept cookies or log in, either by gating content or limiting the user experience for those not logged in.

Additionally, AI is improving the sorting and understanding of text, making more specific contextual targeting possible.

Interest Targeting

An interest group (what user like based on the browsed sites) is stored only in the users's browser and only for up to 30 days. (Google help) Since it "sits" only in the browser, it can be easily blocked by the user.

??Audiences will be smaller.

Topics API

Your browsing activity is recorded and analyzed in your browser, and assigned to interest groups. Just like previously discarded FLoC initiative.

But the main differences are:

• There is a predefined set of interests (topics), currently 350 to pick from, which are human-understandable and transparent. (full taxonomy)
• To maintain privacy, 5% of the time the returned topic is random. (reducing fingerprinting)

PAA - Protected Audiences API (formerly FLEDGE)

PAA set to your browser up to 1,000 interest groups and the primary use is for retargeting.

Bear in mind that “interest group data (such as the ads) can be updated, and an interest group can be enabled for up to 30 days,” (source)

Advertisers can serve ads based on an interest held by the browser but cannot combine that interest with demographic like age.

The Auction

Imagine a sports championship with two groups: contextual and interest group auctions. This is known as Turtledove (Two Uncorrelated Requests, Then Locally-Executed Decision On Victory).

Contextual Auction

The publisher sends a bid request with all available information, and the ad server runs the auction to pick the winner.

Interest Group Auction

The auction includes only the bid floor and the interst group itself. The browser processes the interest groups and picks the highest bidding advertiser. This is complex because each interest group need a separate auctions, and theoretically, any advertiser in the world can participate.

It is unrealistic for a single web browser to handle so much data. Criteo proposed that the auction would be handled by Independent "bookmaker" that would process all the data auctions (SPARROW proposal).

But that raised many questions like who should be this bookmaker. Who will ensure that no personal data leakage.

Dovekey Proposal

So Google replaced the bookmaker with a simple lookup table (Shared Storage API source).

1. Advertisers send bids and creatives for all interest groups they want to enter the auction.
2. The browser sends interest groups to the lookup table.
3. Which checks the bids for the relevant groups.

This heavy lifting is done outside the browser, using a PAA worklet linked to the Shared Storage API.

The Auction Finals

Once the winner from the PAA is known, the final match happens in the browser.

If advertisers set exclusion rules (excluding interest groups) or frequency capping rules, some advertisers may be left out of the game in the final auction that really happens in the browser.

If the winner is from the contextual group, the process is straightforward like today.

Fenced Frames

If the winner is from the interest group, publishers can’t know which audience groups or advertisers won, and advertisers don’t know about the placement. Fenced Frames ensure "blindness" on both sides.

The lookup table (Shared Storage API) can provide the ad creatives. Because the table not only checks who has the highest bid, but also which ad creatives the winners had.

And this is all to show an ad. ?? Here is the final visualization.

Latency Issues

This auction process is complex and takes much longer than current auctions, causing ads to load later, potentially taking an extra 800 to 1,800 milliseconds. - yes over 1 sec! (depending on a source 1500 ms and other test) This means:

1. Viewability Down Good publishers with an average 70% viewability now see averages of 39%. Average publisher with average 50% viewability could drop to 25%.
2. Rise of Passback Ads Publishers already have a "backup" option when the auction takes too long. Publishers will need more of these ads and I think it could lead to discounts for direct deals.
3. Issues with Forecasting and Pacing Publishers might face overbooking issues or no ads at all, potentially leading to ad fraud. (Since even reputable publishers like Forbes use shady techniques already link). Google is working on Private State Tokens to address some of these issues.

??To speed up the process, advertisers should use fewer exclusion criteria and smaller banners (use only 100 kB banners instead of max 150 kB).

Reporting

Because publishers and advertisers get limited data, and the full data is only in the browser.

The reporting will be limited, noisy, delayed and of course aggregated.

Limited

You won't know why your bid didn't win. It will be harder to optimize the future bidding strategy (Was it because of bid limits, audience size or because of latency issues?)

Noisy.

"As the level of detail increases, so does the relative noise in that data." (Private aggregation)

Delayed

Not only is the data delayed, but the viewability measurement is limited. Google claims it works, but IAB flagged way more problems.

Aggregated.

The browser will only send the data aggregated. But it should allow you to get frequency + reach on campaign level and some simple reports.

Conclusion

Audiences (interest groups) will be smaller. Retargeting only up to 30 days.

The auction takes much longer, causing ads to load later and drop in viewable impressions. Use fewer exclusion criteriain targeting, as it can further prolong the auction time.

Lower viewability means less revenue for publishers, who might not want to wait for the auction to finish. This could lead to a rise in passback ads, potentially making direct ad placements more attractive.

Reporting will be limited, noisy, delayed, and aggregated.