Abstract
Google’s "Related Website Sets" (RWS) proposal allows groups of domains to bypass browser privacy protections, such as storage partitioning, if they are deemed "related." This policy rests on the critical assumption that users can accurately identify these organizational relationships, thereby ensuring that data sharing aligns with user expectations. This paper evaluates this assumption through a user study, automated similarity analysis, and a governance audit. Our findings reveal that users fail to recognize related sites nearly 37% of the time. Furthermore, automated tools struggle to link domains by name or visual style, suggesting that RWS may enable unexpected and privacy-harming data sharing behind the user's back.
Introduction
To combat cross-site tracking, modern browsers are increasingly implementing "storage partitioning." This mechanism prevents a website (such as a tracker) from accessing its cookies when it is embedded as a third party on another site. While this is a significant victory for privacy, it can break legitimate multi-domain features, such as a single login session that is intended to work across brand-a.com and brand-b.com owned by the same entity.
Google’s RWS proposal attempts to mitigate this friction by maintaining a list of "related" domains permitted to share storage. The proposal relies on the intuition that if a user understands two sites belong to the same organization, they will not be surprised or harmed when those sites share data. This paper investigates whether that intuition holds up in reality. If users cannot perceive these relationships, RWS effectively creates a "privacy hole" where data is shared without the user's informed consent.
Methods
The authors employed a three-pronged approach to evaluate the viability of the RWS proposal:
- User Study: The researchers recruited 30 participants and presented them with 20 pairs of websites. Participants were tasked with determining if the sites were related by a common company or organization. The study tracked both the accuracy of these determinations and the time taken to reach a decision.
- Automated Similarity Analysis: To determine if the identification problem could be solved algorithmically, the authors measured the similarity of "related" sites using Levenshtein Edit Distance (comparing the text of Second-Level Domains) and HTML Similarity (comparing CSS styles and structural tags).
- Governance Audit: The authors analyzed the official RWS GitHub repository, examining 114 pull requests to understand how the list is managed, the frequency of rejections, and the common technical hurdles for applicants.
Datasets
The study utilized the official RWS list as of March 2024, which contained 41 sets. To provide a baseline, the researchers used the Tranco Top 10K to select random "unrelated" sites for control groups. Website categorization was supported by the Forcepoint ThreatSeeker database to ensure comparisons were made between similar types of sites (e.g., News, Business). Finally, the authors released a new dataset containing 822 generated site pairs and anonymized responses from the user survey.
Analysis of Site Relationships
The "Associated" Loophole
RWS defines different categories for set members. While "Service" sites must have common ownership, "Associated" sites only require a "clear affiliation." The study found that 92.7% of RWS sets include these associated sites, which the data suggests are the most likely to confuse users.
The Effort of Confusion
When participants incorrectly labeled related sites as "unrelated," they spent significantly more time evaluating them—39.4 seconds compared to 28.1 seconds for correct identifications. This suggests that the relationship was not "clearly presented," forcing the user to hunt for clues that were ultimately insufficient.
Naming and Visual Mismatches
Automated tools struggle as much as humans. Half of the associated sites in the RWS list have a Second-Level Domain (SLD) edit distance of 6 or more from their primary domain (e.g., nourishingpursuits.com vs. cafemedia.com), making them nearly impossible to link by name alone. Furthermore, the median "joint similarity" score for HTML structure and style between related sites was a mere 0.04 on a scale of 0 to 1. In many cases, a company’s different brands look nothing alike, further masking their relationship from the user.
Key Results
- High Error Rate: In 36.8% of cases, users failed to identify sites that were actually part of the same Related Website Set.
- Widespread Failure: Nearly 73.3% of all participants made at least one "privacy-harming" error—believing two sites were unrelated when the browser would have treated them as related for data-sharing purposes.
- Governance Friction: The RWS list is difficult to maintain. 58.8% of pull requests to join the list were rejected or closed without being merged, often due to technical complexities like failing to properly host a required
.well-knownJSON file. - Scalability Concerns: While automated checks are fast, the manual review process for RWS sets takes a median of 5 days, raising questions about the system's ability to scale if thousands of companies apply.
Conclusion
The findings suggest that the core assumption of RWS—that users intuitively understand the organizational boundaries of the web—is flawed. If RWS is to move forward without compromising privacy, the browser UI must do more than just assume user knowledge. Future work should investigate how browsers can explicitly signal these relationships to users (e.g., via the address bar) to ensure that cross-site data sharing aligns with actual user expectations. Without such transparency, RWS risks facilitating tracking under the guise of organizational affiliation.
Summary updated