Privacy-enhancing advertising technology did not replace the publisher revenue lost when third-party cookies were removed in this field experiment. The headline estimate is stark: Privacy Sandbox preserved only 4.2% of the revenue lost after cookie removal, while cookie removal reduced publisher revenue by 29.1% in the paper’s abstract. The same research also found higher latency and a 2.9% reduction in impression delivery under Privacy Sandbox. ideas.repec.org
That does not settle the normative question of whether third-party cookies should persist. It does, however, narrow one operational claim that has often been made around privacy technologies: that a browser-mediated alternative can preserve the economic function of third-party cookies for publishers. In the measured setting, the answer was no—not at the observed level of adoption.
For people doing GEO work, this is principally a content-economics result, not an AI-search result. The experiment provides no evidence about whether a site is more likely to be crawled, cited, summarized, or recommended by an AI system. It does provide unusually relevant evidence about the advertising-funded environment in which much open-web content is produced. That distinction matters. A claim that privacy technology will protect publisher revenue is testable here; a claim that it will improve AI visibility is not supported by these sources.
What the experiment actually measured
The researchers partnered with an ad management firm and evaluated more than 200 million ad impressions across more than 5,000 publishers in an open, industry-wide field experiment. ideas.repec.org Tech Xplore describes the experiment as including around 60 million desktop and mobile Chrome users, randomly assigned to cookies-enabled, cookies-disabled, or Privacy Sandbox conditions. Tech Xplore
The design is the important part. This was not a survey of publisher expectations, a modeled estimate based on historical data, or a vendor benchmark. It compared a status quo group with third-party cookies and Privacy Sandbox switched off; a Privacy Sandbox group with cookies off and Sandbox on; and a cookieless group with neither. The experiment arose through a process involving Google and the UK Competition and Markets Authority, according to the researchers’ discussion of the study. MDM by Eric Seufert
That structure makes the result more useful than generic statements that “privacy has a cost.” It identifies two separate comparisons: the cost of losing third-party cookies, and the extent to which a proposed privacy technology restores that loss.
The working-paper abstract reports a 29.1% publisher-revenue reduction when third-party cookies were removed. ideas.repec.org In interview discussion, the authors describe ad prices falling 29% overall and 35% among impressions that carried cookies; they also note that some impressions already lack cookies. MDM by Eric Seufert Tech Xplore reports a figure of about 35% for third-party-cookie removal. Tech Xplore
Those figures should not be casually treated as interchangeable. The supplied material does not fully reconcile the different denominators and outcome descriptions. The safe conclusion is the broad one supported across the sources: cookie removal caused a large adverse publisher-revenue effect in the experiment.
The replacement recovered little, and adoption is part of the story
The central number is not the cookie-loss estimate but the recovery estimate. Privacy Sandbox preserved 4.2% of the lost revenue in the working-paper abstract. ideas.repec.org The authors described the result in the podcast as recovery of only 4% of lost revenue, while Tech Xplore likewise reports about 4%. MDM by Eric Seufert Tech Xplore
There is an important qualification. The researchers say industry adoption of the Privacy Sandbox tools was modest, limiting the amount of revenue recovery one could expect to observe. MDM by Eric Seufert That prevents a stronger claim that privacy-enhancing technologies can never work. It does not rescue the stronger commercial assertion that this implementation, in this experiment, functioned as a revenue-equivalent replacement.
This is a useful standard for evaluating future post-cookie proposals. “Potentially privacy-preserving” and “economically substitutable for cookies” are different propositions. A technology can have a plausible architecture and still underperform in a real auction environment because adoption is incomplete, workflows are immature, or performance introduces new costs.
Latency was not a side issue
The paper’s abstract says Privacy Sandbox increased ad latency and reduced impression delivery by 2.9%. ideas.repec.org In the podcast, the researchers explain the mechanism: Privacy Sandbox moved auction processing onto the user’s device rather than relying on conventional server-side auction processing. They found ad delivery took roughly two seconds under the Sandbox condition compared with roughly one second under the cookie-enabled status quo condition. MDM by Eric Seufert
The authors estimate that about 3% of Sandbox-condition opportunities failed to produce publisher revenue because users navigated away before the ad rendered; they say the problem was worse on smartphones, where computing power is more constrained. MDM by Eric Seufert The abstract’s 2.9% delivery figure and the interview’s approximately 3% explanation are consistent in direction, though presented differently.
For GEO practitioners, this is a reminder not to collapse all web-performance questions into discoverability. A page can have strong source material, structured content, and an effective publishing workflow, yet still depend on an advertising stack whose delivery mechanics affect whether the publisher captures revenue. The research does not show that latency changes AI retrieval or AI citations. It shows that privacy-preserving auction design can introduce a direct economic trade-off for publishers.
The European result makes “one global post-cookie plan” look weak
The podcast discussion reports that, among impressions carrying cookies, cookie removal reduced publisher revenue by roughly 66% for EU traffic and 33% for US traffic. It also reports that only 57% of EU impressions carried cookies, compared with 81% for US users. MDM by Eric Seufert Tech Xplore similarly highlights a roughly 66% EU effect. Tech Xplore
The authors attribute the contrast to the EU’s consent environment: explicit consent requirements produce more opt-outs and make alternative identifiers harder to deploy, leaving fewer alternatives available on European impressions. MDM by Eric Seufert
Yet Privacy Sandbox recovered a larger share of lost EU revenue in the interview results: about 23% in the EU versus 2.4% in the US. The authors offer explanations rather than a definitive causal resolution: Sandbox may matter more where traditional identifiers are weaker, and European firms may have had greater incentives to adopt privacy-enhancing technologies. MDM by Eric Seufert
The practical implication is not that Europe has solved the replacement problem. A 23% recovery share still leaves most of the measured loss unrecovered. It is that implementation and market context change the result materially. Any publisher, platform, or GEO program operating internationally should resist a single blended assumption about identifiers, monetization, and privacy technology performance.
What this does—and does not—mean for GEO
The study is relevant to GEO because GEO is performed on a web shaped by publisher incentives. If advertising revenue falls, publishers may respond through paywalls, subscriptions, or cross-site log-ins; Tech Xplore identifies those as responses to declining cookie use. Tech Xplore But the supplied evidence does not quantify whether those responses improve or worsen the availability of source material for AI systems. It does not measure any effect on AI answer quality, referral traffic from AI products, citation behavior, or brand visibility in generative search.
So the disciplined GEO reading is narrow:
- Do treat the experiment as strong evidence that the publisher-side revenue consequences of cookie loss are material in the studied setting.
- Do treat Privacy Sandbox’s observed revenue recovery as limited, while recognizing that modest adoption constrained the measured outcome.
- Do account for regional differences rather than assuming identifier changes have equal effects across markets.
- Do not use this paper as evidence that cookies, Privacy Sandbox, or any other privacy technology changes AI-search visibility.
The paper was published in Proceedings of the National Academy of Sciences in 2026, according to Tech Xplore, and its research question remains valuable even though Google retired Privacy Sandbox in October 2025, as described by the authors in the podcast. Tech Xplore MDM by Eric Seufert The useful lesson is not that the privacy-versus-revenue debate is finished. It is that replacement claims need field evidence, outcome definitions, regional cuts, and operational performance data—not just a privacy-preserving design on paper.