What Makes a "Bad" Ad? User Perceptions of Problematic Online Advertising

Title of the Paper

What Makes a “Bad” Ad? User Perceptions of Problematic Online Advertising

Bibliographic Information

• Authors: Eric Zeng, Tadayoshi Kohno, Franziska Roesner
• Year: 2021
• Conference: CHI Conference on Human Factors in Computing Systems (CHI ’21)
• Field: Human-Computer Interaction
• Keywords: online advertising, clickbait ads, web ecosystem, user perception, ad content, privacy, dark patterns

Research Background and Issues

• Issues and Challenges: Although online advertising is a crucial part of the modern website economic model, many users express dissatisfaction with its content and presentation. Modern ads often include elements such as "clickbait," "misleading content," or covert promotional tactics, which negatively impact user experience.
• Significance: Advertising is not only a vital component of the internet economy but also a key factor in user browsing experiences. Understanding why users dislike certain ads can provide valuable insights for improving the advertising industry and enhancing user experience.
• Research Motivation and Related Work:
  • The study focuses on issues such as ad intrusiveness, privacy violations, and low-quality content.
  • Previous research lacks a systematic understanding of what users perceive as "bad" or "good" ad content and has not fully leveraged real user perceptions to define advertising problems.
  • This study aims to bridge the gap between user perceptions and ad analysis, providing a theoretical foundation for improving the advertising ecosystem.

Solution

• Proposed Method: The authors conducted user surveys and data analysis to perform qualitative categorization and quantitative research on online ads.
• Innovations:
  1. Developed a taxonomy of user reactions to ad content, including 15 reasons for positive and negative responses.
  2. Collected and labeled 500 ad samples, analyzing 12,972 opinions from 1,000 participants.
  3. Used unsupervised learning techniques to combine user feedback with researcher-provided ad content labels to identify negative ad characteristics.
• Steps and Techniques:
  • Qualitative Research: Gathered user feedback on reasons for liking/disliking ads through surveys.
  • Quantitative Research: Randomly selected 500 ad samples, which were annotated and rated by study participants.
  • Cluster Analysis: Used clustering algorithms based on the distribution of user opinions to classify ads with similar issues.

Research Findings

• Specific Findings:
  • Identified major categories of ads that cause user dissatisfaction, such as "sensational headlines," "low-quality content," "misleading deceptive ads," and "politicized ads."
  • Users expressed predominantly negative attitudes toward the experimental ads, with 45% of ads labeled as "problematic" by more than half of the users.
  • Key reasons for negative reactions included "clickbait," "poor design quality," "lack of clarity," and "politicized content."
• Experimental Results:
  • Over 65% of ads were labeled as "simple"; 20% were considered "clickbait"; 13% involved "poor design quality."
  • Unsupervised learning identified 16 thematic ad clusters, including content farms, political ads, and poorly designed ads.
• Advantages and Limitations:
  • Advantages:
    • Systematically defined "problematic ads" from a user perspective, providing a basis for stakeholders to optimize ad content.
    • Offered a user-labeled ad dataset and classification model, which can be used in future automated ad detection research.
  • Limitations:
    • The study population primarily consisted of younger users and ad blocker users, potentially underestimating positive evaluations of ads.
    • Focused only on specific types of ads (third-party programmatic ads), excluding social media and video ads.

Limitations and Future Directions

• Limitations:
  • The study did not cover all ad formats or perspectives from global users.
  • The selected ad samples represent a snapshot at a specific point in time, which may not fully reflect the broader advertising ecosystem.
  • User understanding of the taxonomy may vary, affecting annotation consistency.
• Future Directions:
  • Measuring post-click "regret": Investigating which ads waste users' time and attention resources.
  • Analyzing target audiences of problematic ads: Exploring whether ads are targeted at specific vulnerable groups.
  • Developing classifiers for automated detection of "problematic ads": Advancing advertising ecosystem governance based on user-labeled distribution learning.
  • Improving ad content policies: Suggesting the inclusion of user feedback mechanisms to enhance transparency and quality in ad content.

Conclusion

This study proposes a user-centered approach to systematically understand the "good" and "bad" characteristics of online ads. The analysis of ad design and content issues provides a theoretical foundation for improving the overall quality of online advertising and user experience. It also highlights directions for collaboration among policymakers, ad platforms, and technical researchers.