TL;DR
This paper presents the first large-scale analysis of how Facebook users are exposed to online advertising, revealing that users see a median of 70 ads per week, which constitute 10-15% of their newsfeed content. Surprisingly, Facebook appears to overprofile users, with a significant portion of assigned interests being irrelevant to the ads they receive, and engagement with ads grows logarithmically with impressions, suggesting frequency caps are crucial.
What Problem Does This Paper Address?
Online advertising is a massive industry, forming the backbone of revenue for many internet services like Google and Facebook. While much research focuses on optimizing ad engagement, understanding privacy implications, and improving ad delivery, there's a significant gap in understanding the actual, large-scale exposure of real users to online advertising. Previous studies often relied on bots or limited datasets. This paper aims to fill that gap by providing a comprehensive, user-centric analysis of Facebook's advertising ecosystem, quantifying user exposure to ads and advertisers, and investigating the effectiveness of Facebook's user profiling for ad targeting.
Methods
The researchers utilized a custom web-browser plugin called the Data Valuation Tool (FDVT) to collect data from over 5,000 real Facebook users. This tool monitored ads delivered to users, their positions (newsfeed vs. right-side), and the landing page URLs. To identify advertisers, they extracted web domains from these URLs and categorized them using McAfee's services. They employed statistical analysis to quantify ad exposure over various time windows (session, day, week, month), analyze the distribution of ads across advertisers and categories, and assess user profiling by measuring the semantic similarity between ad categories and user interests. Engagement was measured by click-through rates (CTR) and analyzed in relation to ad impressions.
Datasets
The study used a large dataset collected between October 2016 and May 2018, comprising over 7 million ads from 140,000 unique advertisers delivered to more than 5,000 FDVT users across nearly 600,000 sessions. This dataset is noted as the largest used in a study analyzing user exposure to online advertising.
Key Insights and Ideas
- Ad Saturation: Users are not just seeing a few ads; they are consistently exposed to a significant volume, making ads a substantial part of their online experience.
- Overprofiling is Real: Facebook's sophisticated profiling system, while intended to deliver relevant ads, seems to be assigning users a vast number of interests, many of which are irrelevant to the ads they actually see. This suggests a potential inefficiency and privacy concern.
- Diminishing Returns on Impressions: Bombarding users with ads doesn't necessarily lead to proportionally higher engagement. The benefit of showing more ads to a user decreases logarithmically, highlighting the importance of optimizing frequency caps.
- First Impressions Matter (Slightly): While clicks are somewhat evenly distributed across multiple ad impressions, the very first impression holds a slightly higher probability of engagement.
- Advertiser Concentration: A small number of advertisers are responsible for a large portion of the ads delivered to users, indicating a concentrated advertising market on Facebook.
Key Results
- Ad Volume: Facebook users are exposed to a median of 70 ads per week, 6 ads per session, and 0.8 ads per minute.
- Newsfeed Content: Ads constitute 10-15% of the content in users' newsfeeds. A 1% increase in this proportion could translate to an estimated $8.17 million weekly revenue increase for Facebook.
- Advertiser Reach: On average, users are targeted by 4.5 advertisers per session, 12 per day, 33 per week, and 67 per month.
- User Profiling Inaccuracy: In the best-case scenario, only 22.76% of the interests Facebook assigns to users are actually related to the ads they receive.
- Engagement Growth: The probability of an advertiser getting a click grows logarithmically with the number of ad impressions, with the first impression having a slightly higher engagement probability.
- User Segmentation: While most users receive a similar proportion of ads, 8.37% of users are systematically exposed to a higher volume of ads.
Limitations and Future Work
The authors acknowledge that the FDVT users might not be a perfectly representative sample of all Facebook users. They also note that identifying advertisers solely through landing page domains can be challenging, and some URLs were no longer active or led to intermediary domains.
Future work suggested includes further investigating the role of Facebook users as "products" being auctioned, replicating experiments in different time windows to observe long-term trends and potential differences due to evolving regulations (like GDPR), and exploring the optimization of frequency caps for advertisers to maximize ROI.
Summary updated