Lessons from the Adoption and Deprecation of the Privacy Sandbox Web APIs
Source: arXiv:2606.26390 · Published 2026-06-24 · By Yohan Beugin, Paul Barford, Patrick McDaniel
TL;DR
This paper presents a comprehensive longitudinal measurement and analysis of the Google Privacy Sandbox web APIs, introduced in 2019 and deprecated in late 2025. The Privacy Sandbox aimed to provide privacy-preserving alternatives to third-party cookies (TPC) with new browser APIs supporting advertising, conversion measurement, and fraud detection use cases. By analyzing HTTP Archive crawls of the top 100k websites and aggregated Chrome telemetry data over a 5+ year period, the authors characterize adoption trends and usage patterns across the ecosystem. They find that adoption was limited and highly uneven with only a few web actors integrating very specific APIs, often in inconsistent ways, and at low volumes compared to legacy TPC-based techniques. The paper also highlights ongoing privacy risks such as browser fingerprinting through User-Agent Client Hints and the lack of widespread enforcement or default privacy protections in Chrome, unlike other browsers. The study yields actionable recommendations for future web privacy proposals, emphasizing precise evaluation, clear use cases, non-exploitable APIs, and default deployment of protections.
Key findings
- F1: Only 5 APIs ever surpassed 11.7% of page loads with detected usage; majority of APIs showed very low or almost no adoption (Fig 2).
- F2: API adoption increased following API introductions in 2023 but peaked and stagnated by Q3 2024, coinciding with Google postponing and then canceling TPC deprecation (Figures 7, 8, 9).
- F3: Majority of adopters were present from early days, with 295 attestations and 73 Related Website Sets showing persistent engagement over years.
- F4: Most supported APIs remained little used, with usage rates ≤5.2% page loads.
- F5: CHIPS API adoption covered ~34.5% of page loads but only accounted for 6.7% of TPC accesses on average.
- F6: UA Client Hints API was invoked on over 62% of page loads and found to be abused for browser fingerprinting, weakening UA reduction efforts.
- Additional Finding: Tracking and TPC limitations in Chrome remain mostly opt-in, while other browsers (Brave, Firefox, Safari) have enabled these protections by default.
- Finding from HTTP Archive: Even on top 100k sites, API adoption was sporadic and limited to a small subset of domains.
Threat model
The threat model considers the web ecosystem actors including advertisers, publishers, browsers, and users. The adversary is primarily the browser or third-party trackers attempting cross-site tracking through mechanisms like third-party cookies or fingerprinting. The study assumes adversaries seek to exploit browser APIs to track users or measure conversions without consent or user privacy guarantees. However, it does not consider active adversarial manipulation or attacker-driven evasion of measurements but focuses on ecosystem adoption patterns and privacy risks. It implicitly assumes that browser vendors can enforce or disable APIs and that sites can choose to adopt or not. The threat excludes active network attackers or browser vulnerabilities and focuses on privacy policy and API design limitations.
Methodology — deep read
The study focused on a threat model of web tracking ecosystem actors including advertisers, publishers, and browser vendors, evaluating ecosystem response rather than direct adversarial attacks. Data provenance included two primary sources: (1) Chrome Telemetry, collecting feature usage telemetry aggregated daily across millions of Chrome users globally from January 2020 through April 2026, covering stable, beta, and dev channels across platforms; (2) HTTP Archive monthly crawls of the CrUX top 100k websites from March 2021 to April 2026, recording HTTP response headers, scripts, and page content under a controlled, stateless, non-interactive crawling setup.
The authors instrumented detection of 13 primary Privacy Sandbox APIs through respective Blink feature flags, permissions policies, and API invocation signatures mapped from official developer docs and explainer documents. This included monitoring JavaScript method calls, HTTP headers, and feature usage to estimate presence and activation of APIs during real user browsing (Chrome Telemetry) and website crawling (HTTP Archive).
To reduce false positives and cost in HTTP Archive analysis, they filtered pages by prior Blink feature detections, then ran syntactic matching on root page content and related response headers to confirm API usage instances. They also incorporated data from public chromium source code and the Google Related Website Sets repository to identify attested API adopters.
The 13 APIs ranged from Federated Credential Management (FedCM), Attribution Reporting, CHIPS, Topics, UA Client Hints, to storage access mechanisms like Shared Storage and Storage Access.
Evaluation involved longitudinal trend analyses of adoption rates (% page loads with API usage detected) over 5+ years, with comparison points anchored to major project milestones such as TPC deprecation delays and API general availability dates. They analyzed early versus late adopters, consistency over time, and relative usage compared to TPC. Privacy risks were examined qualitatively where APIs weakened privacy guarantees or introduced unmitigated fingerprinting risks.
The entire data collection and analysis pipeline has been open-sourced as a reproducible artifact, encouraging further research and extensions. However, exact replication depends on ongoing access to Chrome Telemetry and HTTP Archive datasets which are public but large and complex.
Technical innovations
- A large-scale, longitudinal measurement methodology combining real user telemetry with large web crawls to track adoption of browser privacy APIs in-the-wild over multiple years.
- A practical filtering heuristics approach using Blink feature detections to guide focused exhaustive syntactic matching in massive HTTP Archive datasets to detect API usage with manageable costs.
- Cross-referencing multiple data sources including Chromium source control and attestation lists to identify and characterize active API adopters and usage patterns precisely.
- Systematic alignment of adoption metrics with project milestone events to interpret ecosystem response dynamics in real time.
- Highlighting fingerprinting risks posed by UA Client Hints through telemetry-based prevalence and abuse analysis.
Datasets
- Chrome Telemetry — millions of Chrome user browsing sessions aggregated daily — public aggregated from Google
- HTTP Archive — monthly crawls of top 100k websites from desktop and emulated mobile clients — public
- Chromium source tree and Google RWS repository — version controlled attestation and related website sets data — public open-source
Baselines vs proposed
- General TPC tracking: continues to be supported and prevalent in Chrome while Privacy Sandbox adoption remains opt-in or low.
- API usage rates: Top 5 APIs at max 11.7% page loads vs most others <5.2%.
- CHIPS API: coverage ~34.5% of page loads but only 6.7% mean TPC usage reduction.
- UA Client Hints: >62% page loads API call rate vs browser UA reduction effort weakened.
Limitations
- Analysis focuses on adoption measurement and usage prevalence; no direct adversarial evaluation or attack simulation performed.
- HTTP Archive crawler is non-interactive and stateless, potentially underestimating API usage that requires user interaction or logged-in state.
- APIs usage detection via syntactic matching approximate—may include false positives/negatives depending on code analysis heuristics.
- Chrome Telemetry data derived from users who opted in to telemetry, potentially biasing demographic or usage representation.
- Study limited to publicly available data sources without access to Google’s internal telemetry or confidential logs.
- No in-depth user privacy impact measurement beyond aggregate API usage and fingerprinting risk qualitative discussion.
- The privacy and legal status of APIs like UA Client Hints remains unsettled; enforcement and policy contexts could affect adoption independently.
Open questions / follow-ons
- What incentive structures or regulatory pressures could have realistically increased broader ecosystem adoption of Privacy Sandbox APIs?
- How might future browser privacy proposals be designed to better prevent abuse scenarios such as fingerprinting and unauthorized data leakage?
- What empirical methods can be developed to more precisely evaluate end-user privacy impacts of emerging web privacy APIs under real-world usage?
- How can browser vendors coordinate more effectively to align privacy defaults and enforcement mechanisms globally to avoid fragmentation?
Why it matters for bot defense
For bot-defense and CAPTCHA practitioners, this paper highlights critical lessons about designing and deploying privacy-enhancing web APIs in adversarial environments. The limited, uneven adoption and eventual rollback of Privacy Sandbox APIs illustrate challenges of balancing utility and privacy within the complex incentive structures of the web ecosystem. Bot-defense engineers should note that relying predominantly on "fix it in the browser" approaches risks fragmentation, opt-in usage, and slow adoption that can undermine security gains. The findings underscore the importance of clear communication of API use cases, strong default privacy protections, and monitoring for side-channel risks like fingerprinting through APIs such as UA Client Hints. Furthermore, this research suggests the value of platforms adopting privacy-by-default settings to reduce attackers’ available signals, including bots leveraging covert browser channels or storage access. As CAPTCHA design increasingly integrates with browser APIs (e.g., Private State Tokens for fraud prevention), practitioners must be mindful of real-world adoption trends and how ecosystem incentives shape deployment, to avoid building on fragile or abandoned foundations. Overall, the longitudinal measurement approach and open datasets also provide methodology references for future bot-defense studies monitoring web-level API usage and behavioral signals relevant to bot detection.
Cite
@article{arxiv2606_26390,
title={ Lessons from the Adoption and Deprecation of the Privacy Sandbox Web APIs },
author={ Yohan Beugin and Paul Barford and Patrick McDaniel },
journal={arXiv preprint arXiv:2606.26390},
year={ 2026 },
url={https://arxiv.org/abs/2606.26390}
}