How Product Recommendations Influence Purchase Decisions

Abstract

Product recommendations are ubiquitous in our online experience, guiding us towards products we might like on Amazon, Spotify, or even news websites. But have you ever wondered how these recommendations work? This blog post will help you understand how these “recommender systems” shape our purchasing decisions and how marketers can leverage this knowledge for better results.

Findings

Recommender systems aim to help customers discover products that match their interests. They rely on information gathered from past customer behavior, analyzing what they’ve purchased, listened to, or read. There are two main approaches to generating recommendations:

  • Content-Based Recommendations: These systems focus on similarities between items. For instance, if you’ve watched numerous comedies, you’ll likely receive more recommendations within that genre. Think of it like finding something similar to what you’ve already liked.
    • Companies use data about items to create “item profiles,” which include details like actors, directors, or genres. [1, 5]
    • For example, a movie profile might include its actors, director, year of release, and genre. [1]
    • These profiles are then used to construct “user profiles,” essentially a record of the things you enjoy. [1, 5]
    • For instance, if you frequently watch movies with Julia Roberts, your user profile will reflect that preference. [1]
    • By comparing your user profile with item profiles, companies can suggest items similar to those you’ve shown interest in.
  • Collaborative Filtering Recommendations: This approach focuses on similarities between people rather than products. It analyzes what similar individuals have liked to predict what you might enjoy.
    • Companies use a “utility matrix” to track your ratings or purchase history. [1, 5]
    • For example, a utility matrix might show that you rated Harry Potter movies highly, while another user gave Star Wars movies high ratings. [1]
    • Companies use various methods to calculate the similarity between you and other customers, such as analyzing the types of movies or music you enjoy and comparing it to others’ preferences. [1, 5]
    • One method used to measure similarity is the “cosine distance,” which calculates how close the directions of two vectors are, with vectors representing the ratings of different users or items. [1]
    • These similarities can then be used to predict your liking for items you haven’t interacted with before.

Key Takeaways

  • The Power of the Long Tail: Online stores offer a vast array of items, but physical stores have limited space. This means online stores must use recommendation systems to help customers discover a wider variety of products. The “long tail” represents the vast amount of less popular items available online but rarely found in physical stores. For example, a physical bookstore might only have a few thousand books on its shelves, while Amazon offers millions. [1]
  • User-Based Framing Can Be More Effective: Research has shown that recommendations framed as “people who like this also like…” can be more effective in encouraging clicks, compared to recommendations framed as “similar to this item”. [2] This is because user-based framing provides additional information about shared tastes, which helps customers predict their own liking of the recommended item.
  • Experience Matters: Customers with more experience in a specific product category are less likely to trust recommendations based on shared tastes. For example, someone who has been listening to music for a long time might not be swayed by recommendations based on other users’ preferences. This is because experienced customers often have more refined tastes and might not see a shared interest in one or two items as a sign of true similarity. [2]
  • Attractiveness Matters: Recommendations are more likely to be clicked on if they suggest an item that’s similar to one the customer already likes. For example, if you’ve already rated a movie highly, you are more likely to click on a recommendation for a similar movie. This is because customers learn about their preferences through their reactions to different products, and attractive items serve as a strong indicator of personal taste. [2]
  • Avoid Dissimilarity Cues: It’s crucial to be mindful of cues that suggest the customer is different from other users who liked the recommended product. For example, if a recommendation suggests a product that’s popular with teens, and the customer is an adult, they may be less likely to trust the recommendation. This is because customers tend to avoid dissimilar others’ tastes, and cues that suggest a mismatch in preferences can undermine the trust in a recommendation. [2]
  • Recommender Systems Are Sensitive to Data Perturbations: Even a tiny change to the data used to train a recommender system can drastically change the recommendations generated for all users. This is because changes in the training data affect how the system learns about user preferences, which can lead to unexpected shifts in the recommendations. [3, 4]
  • Low-Accuracy Users Are More Susceptible: Users who are already receiving low-quality recommendations are more likely to be affected by changes in the training data. For example, if the recommender system has a hard time predicting what a specific user likes, small changes in the training data can cause much larger changes in their recommendations. [3]
  • Recommendation Systems Are Not Just About Accuracy: The accuracy of a recommendation system, how well it predicts what a customer will like, is not the only thing that matters. You also need to consider how stable it is. A stable system will generate similar recommendations even when the training data is slightly changed, while an unstable system can generate drastically different recommendations based on small changes. [3]
  • Recommender Systems Are Often Used in Unexpected Ways: For example, in a fascinating story, Amazon’s recommender system helped to boost sales of an older book called Touching the Void after it recognized that many people who were buying Into Thin Air (another book on mountain climbing) were also interested in Touching the Void. Without online booksellers and their recommendation systems, the less popular book may not have found many readers. [1]
  • Recommender Systems Can Be Improved: The Netflix prize, a competition that offered a million dollar reward for the best algorithm to improve movie recommendations, highlighted the ongoing efforts to improve these systems. This competition led to the development of new algorithms that were able to better predict user preferences, demonstrating the constant innovation in this field. [1]

Conclusion

Product recommendations are a powerful tool for marketers. By understanding how these recommendations work, you can create more effective marketing strategies that help customers discover the products they’ll love. As you develop your strategies, remember the importance of factors like user experience, product attractiveness, and the potential for instability. By considering these factors, you can create recommendations that are more likely to be successful and drive sales.

Sources

  1. Ullman, Jeffrey D. Mining of Massive Datasets. Cambridge University Press, 2011. http://infolab.stanford.edu/~ullman/mmds/ch9.pdf
  2. Gai, Phyliss Jia, and Anne-Kathrin Klesse. “Making Recommendations More Effective Through Framings: Impacts of User- Versus Item-Based Framings on Recommendation Click-Throughs.” Journal of Marketing, vol. 83, no. 6, 2019, pp. 61–75. https://journals.sagepub.com/doi/pdf/10.1177/0022242919873901
  3. Oh, Sejoon, et al. “Rank List Sensitivity of Recommender Systems to Interaction Perturbations.” Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM ’22), 2022. https://faculty.cc.gatech.edu/~srijan/pubs/CASPER-sejoon-CIKM-2022.pdf
  4. Ayday, Erman, et al. “A Recommender System Based on Belief Propagation over Pairwise Markov Random Fields.” 2012 50th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2012, pp. 1265–1272. https://fekri.ece.gatech.edu/BP_publications/allerton_2012.pdf
  5. Bodapati, Anand V. “Recommendation Systems with Purchase Data.” Journal of Marketing Research, vol. 45, no. 1, 2008, pp. 77–93. https://www.anderson.ucla.edu/faculty/anand.bodapati/Recommendation-Systems-with-Purchase-Data.pdf

What is the best quiz for you business?

Quizzes are super effective for lead generation and selling products. Find the best quiz for your business by answering a few questions.

Take the quiz