💻 computer science

Auditing the Impact of Cross-Site Web Tracking on YouTube Political and Misinformation Recommendations

This paper introduces a sock-puppet-based experimental framework to audit how off-platform cross-site tracking by Google influences YouTube's political and misinformation recommendations, while evaluating the effectiveness of privacy-focused browsers in mitigating such tracking-driven content bubbles.

Original authors: Salim Chouaki, Savaiz Nazir, Sandra Siby

Published 2026-03-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Salim Chouaki, Savaiz Nazir, Sandra Siby

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your digital life is like a personal tour guide who knows everything you do. You might think this guide only watches what you do inside a specific building (like YouTube), but in reality, this guide is also peeking through the windows of every other building you visit on the internet (like news websites).

This paper is about investigating whether that "peeking" changes what the tour guide shows you when you return to the main building.

Here is the breakdown of the research in simple terms:

1. The Big Question: Does What You Read Elsewhere Change Your YouTube?

Most people know that if you watch a lot of conspiracy theory videos on YouTube, YouTube will show you more of them. This is called a "filter bubble."

But this study asks a different question: What happens if you read a fake news article on a regular website, but you never watch a video about it on YouTube?

Does Google (which owns YouTube) use its "spy tools" (trackers) to see that you read the article, and then secretly change your YouTube homepage to show you videos that match that fake news?

2. The Experiment: The "Sock Puppet" Spies

To find out, the researchers couldn't just use real people (because real people have messy habits). Instead, they created 100 "Sock Puppets."

Think of these sock puppets as robot clones with fresh, clean slates.

The Setup: They gave each robot a brand-new browser, like a new car with no history.
The Warm-up: To get the robots started, they had them watch one boring sports video. This was just to wake up YouTube's recommendation engine so it would start showing them things.
The Test: The researchers split the robots into different groups:
- Some robots visited Left-wing news sites.
- Some visited Right-wing news sites.
- Some visited Extremist sites.
- Some visited sites full of Fake News/Misinformation.
- One group (the control) visited nothing.

While visiting these sites, the robots acted like normal humans: they scrolled, clicked "I agree" to privacy banners (so the trackers could follow them), and stayed for a minute.

3. The Two Worlds: The "Glass House" vs. The "Fortress"

The researchers ran this experiment in two different types of browsers to see if privacy tools matter:

World A (The Glass House): This is a standard browser (like Chrome) with no privacy shields. It's like walking through a city with a glass house; everyone can see exactly where you go and what you look at.
World B (The Fortress): This is a privacy-focused browser (like Brave) with "Shields Up." It's like walking through the city wearing a cloak and a mask; the trackers can't see where you go.

4. What They Found (The "Aha!" Moment)

After the robots visited the news sites, the researchers checked their YouTube homepages.

In the Glass House (No Privacy): The robots that visited fake news sites started seeing more videos related to those fake claims on their YouTube homepage. Even though they never watched those videos, the "tour guide" had seen them reading the articles and thought, "Oh, they like this stuff, let's show them more!"
In the Fortress (With Privacy): The robots that visited the same fake news sites did not see a change in their YouTube recommendations. Because the browser blocked the trackers, the tour guide didn't know they had visited those sites, so the recommendations stayed neutral.

5. Why This Matters

This study proves that YouTube's algorithm isn't just looking at what you watch on YouTube; it's looking at what you read everywhere else on the web.

The Danger: If you read a conspiracy theory on a news site, YouTube might start feeding you conspiracy videos, creating a "bubble" even if you never clicked on a video about it.
The Solution: The study shows that using privacy-focused browsers (the "Fortress") can actually stop this from happening. It acts like a shield, preventing the algorithm from building a bubble based on your off-platform reading habits.

The Bottom Line

Think of your internet experience as a garden.

Without privacy: The gardener (Google) sees you picking a poisonous flower in a neighbor's garden and immediately plants more poisonous flowers in your own yard.
With privacy: The gardener can't see what you did in the neighbor's garden, so your yard stays safe and balanced.

The researchers are now planning to dig deeper to see how long these effects last and how they change if you are logged into a Google account versus using an anonymous one. But for now, the message is clear: Your browsing history outside of YouTube is shaping what you see inside it.

1. Problem Statement

While YouTube is a primary news source for many users, existing research on its recommendation algorithm suffers from a "closed-platform" limitation. Prior studies have primarily analyzed how recommendations evolve based on on-platform watch history (e.g., watching political videos on YouTube).

However, Google's privacy policy and technical infrastructure indicate that the recommendation system also utilizes off-platform browsing data. Google collects user activity from third-party sites via trackers (e.g., Google Analytics, embedded YouTube players) to personalize content across its ecosystem.

The Gap: No prior work has systematically measured how visiting news articles on external websites influences the political bias or misinformation content of YouTube recommendations.
The Risk: With Google trackers present on 72% of sites and 45% of all websites, off-platform browsing could create "tracking-driven filter bubbles," exposing users to polarized or false content even without watching such content on YouTube itself.

2. Methodology

The authors propose a sock-puppet-based experimental framework to audit the platform. The study uses automated browser instances to simulate user behavior across three distinct phases:

A. Experimental Setup

Subjects: Automated "sock puppets" (browser instances) with fresh, cleared profiles.
Initialization: To bypass YouTube's requirement for watch history before showing recommendations, puppets first watch a neutral sports video to generate a baseline.
Environments: Experiments are run in two distinct browser configurations to test the impact of privacy:
1. Tracking-Permissive: Google Chrome (default settings, third-party cookies allowed).
2. Tracking-Rstrictive: Brave browser (Shields up, third-party cookies and tracking requests blocked).
Account Status: All puppets remain logged out of Google accounts to isolate the effect of cookie-based tracking from account-based data.

B. Exposure Phase (Independent Variable)

Six groups of sock puppets are exposed to different types of external content:

Political Ideologies: Four groups visit articles from outlets classified by Media Bias/Fact Check (MBFC) as Extreme-Left, Left, Right, and Extreme-Right.
- Data Source: 50 outlets per ideology $\times$ 20 articles = 1,000 articles per group.
Misinformation: One group visits articles tagged as misinformation by PolitiFact (2020–2025).
- Data Source: 52,000 articles; 20 randomly sampled per puppet.
Control: One group visits no news URLs.

Behavior Simulation: Puppets accept consent banners (to enable tracking), scroll randomly, and dwell on pages for up to one minute.

C. Measurement Phase (Dependent Variable)

After exposure, puppets return to the YouTube homepage to collect recommended videos.
Duration: The cycle (Exposure $\to$ Measurement) is repeated daily for five consecutive days to observe temporal persistence.
Analysis Technique:
- Misinformation Detection: Video titles and transcripts are compared against a corpus of PolitiFact false claims (2020–2025) using MPNet (a pre-trained transformer model).
- Metric: Cosine similarity between video embeddings and misinformation claim embeddings is calculated to quantify the "misinformation score."

3. Key Contributions

First Cross-Platform Audit: This is the first study to explicitly link off-platform browsing activity (news consumption) to on-platform (YouTube) recommendation shifts, moving beyond the closed-platform view.
Privacy-Centric Design: By comparing tracking-permissive and tracking-restrictive environments, the study isolates the specific causal role of cross-site tracking in content curation.
Automated Framework: The development of a scalable sock-puppet framework that simulates realistic user browsing patterns (consent acceptance, scrolling, dwell time) to trigger tracking mechanisms.

4. Preliminary Results

The authors deployed 100 sock puppets, collecting over 51,000 recommended videos (23,854 pre-training, 27,162 post-training).

Misinformation Exposure:
- Tracking-Permissive (Chrome): After visiting misinformation articles, the cosine similarity between recommended videos and known false claims increased by +0.0134. This indicates that off-platform visits directly influenced YouTube to recommend content aligned with those false claims.
- Tracking-Restrictive (Brave): No significant increase in similarity was observed. The recommendations remained stable despite the off-platform visits.
Implication: Cross-site tracking is a viable mechanism for injecting misinformation into user feeds. Privacy-focused browsers effectively neutralize this specific vector of influence.

5. Significance and Future Work

Significance: The study challenges the notion that "filter bubbles" are solely a result of on-platform behavior. It demonstrates that off-platform data flows are a critical, previously unmeasured factor in political polarization and misinformation spread. It validates that privacy tools (like Brave) can act as a defense against algorithmic manipulation driven by cross-site tracking.
Future Directions:
- Political Bias Analysis: Extending the analysis to measure shifts in political ideology (Left/Right) rather than just misinformation.
- Temporal Dynamics: Investigating how quickly recommendations change after exposure and how long the effect persists.
- Logged-in vs. Logged-out: Comparing the impact of tracking on anonymous users versus those signed into Google accounts to determine if account data amplifies these effects.

Conclusion

The paper provides empirical evidence that cross-site web tracking significantly alters YouTube's recommendation landscape, specifically increasing exposure to misinformation when users browse external news sites. Conversely, blocking third-party trackers effectively prevents these algorithmic shifts, highlighting the importance of browser-level privacy protections in mitigating the spread of political polarization and misinformation.