TL;DR
This paper introduces a novel graph-based methodology to uncover website administration by analyzing Google's publisher-specific IDs. The research reveals that while most websites are managed by a single entity, a growing number of intermediary publishers are emerging, and websites are increasingly shifting from large administrators to smaller ones over time.
What Problem Does This Paper Address?
The digital advertising ecosystem is incredibly complex, making it difficult to answer fundamental questions like "Who owns and monetizes which websites?" and "How has this administration landscape evolved?" This opacity hinders transparency and can obscure the influence of various entities on online content. Existing research has touched upon aspects of this problem, but a systematic, large-scale study to map out website administration using readily available identifiers has been lacking.
Methods
The core of this work is a graph-based methodology. The researchers first identify Google publisher-specific IDs (like Google AdSense Publisher IDs, Google Tag Manager Container IDs, and Google Analytics IDs) embedded within the top 1 million websites. They then construct two types of graphs:
- Bipartite Graphs: These graphs represent relationships between websites and the specific IDs they use. An edge exists from a website node to an ID node if that website contains the ID.
- Metagraphs: These graphs represent relationships directly between websites. An undirected edge connects two websites if they share a common publisher-specific ID. The weight of this edge increases with the number of shared IDs, indicating a higher confidence that the websites are administered by the same entity.
To validate their metagraph, they employ community detection algorithms (specifically Girvan-Newman) to identify clusters of websites likely managed by the same administrator. They also perform historical analysis by crawling websites over a two-year period to track changes in administration.
Datasets
The primary dataset used is the Tranco list, which ranks the top 1 million websites. The researchers also utilized data from HTTPArchive for their historical analysis, crawling up to 8 million websites across multiple snapshots. For website categorization, they leveraged data from SimilarWeb.
Key Insights and Ideas
- Dominance of Single Administrators: The vast majority (around 90%) of websites are associated with a single publisher ID, suggesting a straightforward administration model for most.
- Rise of Intermediaries: A significant finding is the emergence of "intermediary publishers" – third-party services that help content creators monetize their sites. These entities are increasingly managing large portfolios of websites.
- Decentralization Trend: Over time, there's a notable trend of websites moving away from large, established administrators towards smaller ones.
- Publisher Size and Popularity Correlation: Smaller publishers tend to manage less popular websites, while larger publishers often manage more popular ones, creating a "rich get richer" dynamic.
- Category Preference: Publishers often exhibit a preference for managing websites within similar categories, indicating a non-random approach to portfolio expansion.
- Potential for Malicious Activity Detection: Websites with an unusually high number of publisher-specific IDs were found to sometimes be associated with malicious activity, suggesting this methodology could be a tool for identifying such sites.
Key Results
- Approximately 90% of websites are associated with a single publisher ID.
- 340 publishers were found to monetize traffic from more than 10 websites each.
- The top 10 publishers in their dataset were found across over 4,200 websites.
- Historical analysis showed that "mega-publishers" (managing >100 websites) lost approximately 25% of their population over a two-year period, while small publishers saw significant growth.
- The Girvan-Newman algorithm, after edge pruning, provided better community detection in 42.5% of cases compared to another method.
- News and Media was the most prevalent website category among those analyzed (24.5%).
Limitations and Future Work
- False Positives: The methodology relies on regular expressions to detect IDs, which could lead to false positives if alphanumeric strings coincidentally match the patterns without being actual identifiers. While filtering techniques are used, this remains a potential issue.
- Google-Centric: The study focuses exclusively on Google's publisher-specific IDs. A more comprehensive view would require analyzing identifiers from other major ad networks and analytics providers.
- Category Data Bias: The reliance on SimilarWeb for website categorization introduces potential biases or inaccuracies.
- Future Work: The authors suggest further investigation into the outlier websites with numerous publisher-specific IDs, as this may indicate anomalous or malicious behavior. Expanding the analysis to include identifiers from other services is also a clear future direction.
Summary updated