The Perils of Personalization
Learnings from Tracy Anderson to Glassdoor Recommendations
I had my phone in my pocket as I chatted with friends this past Memorial Day weekend. Somehow we landed on the topic of how unrelatable Gwyneth Paltrow is, which then led to how absurd it is that her BFF Tracy Anderson tells women not to lift any weights that are more than 3 lbs. Lo and behold, who shows up in my Facebook ads the next day? Tracy Anderson!!!
I never took any action to engage with Tracy Anderson beyond that casual conversation. I never searched for Tracy Anderson. In fact, I never even took my phone out of my pocket during the conversation. Conclusion? My phone was spying on me, and this kind of personalization is both commonplace and creepy.
When you design products, how do you find the balance between intruding on user privacy and providing user value? The answer lies in two key parameters: expectation and intent.
User Privacy vs. User Value Decision Tree
In the Tracy Anderson example, I didn’t give explicit signals of interest. My expectation was that I was not being tracked, and I was not looking to work out with Tracy. As a result, I was offended by the intrusion on my privacy because I felt that the personalization served the company, not me. If your answers are “No” and “Company”, it’s best to reconsider your approach to avoid this peril of personalization.
It’s true that another user might perceive the same Tracy Anderson ad differently. What if a group of friends were having a conversation about how effective Tracy’s workouts are and then the ad showed up with a discount. In this case, the ad would be appreciated and serve the user, whereas in my case, the ad served the company.
Although the execution is the same, its reception can be different because intent can be hard to discern across all users. So even when personalization seems to serve the user, you should still tread carefully whenever data is being implicitly tracked. If your answers are “No” and “User”, you can use personalization but proceed with caution.
Example: Tracy Anderson ad
Contrast the Tracy Anderson example with my curiosity about Goop’s This Smells Like My Vagina Candle. The concept of the candle sounded like a satire from the Onion, so naturally I had to look at the candle’s product details page to confirm if it was really inspired by the scent of Gwyneth’s lady parts. No more than fifteen minutes later, I received an email asking me if I was “Looking for Something?” and encouraging me to “Make it yours”. I passed on the candle and immediately unsubscribed from the email, but was otherwise unperturbed by this incident. I understood that I gave an explicit signal by looking at the page, so even though I had not given Goop my email address nor put it in my cart, this personalization felt far less creepy. If your answer is “Yes”, you can go ahead with personalization.
Example: Goop
The knowledge that my data is being tracked sets expectations so that I don’t react to retargeting. But it also sets expectations that if you serve a recommendation to me, it better be good.
Imagine my surprise upon getting an email from Glassdoor Recommendations with the subject line, “You look like a good fit for the job at FedEx Office.” I expect Glassdoor to have lots of data about me. I’ve looked at Glassdoor reviews to research what employees have to say about a company. I’ve looked at what types of questions they ask in product manager interviews. I’ve looked at salary ranges for product manager roles. And yet, the job emails that I get from them seem completely random. Glassdoor has recommended that I apply for jobs as a retail associate at FedEx one day and a litigation associate the next day.
The sad truth is that a couple years ago, the recommendations actually seemed better than they are today. I’m guessing that sometime in the past year, Glassdoor updated the machine learning model that powers these recommendations. After the algorithm changed, I can no longer explain why these jobs were recommended for me or even find a common thread behind the jobs. Whether it’s bad data or a busted algorithm, Glassdoor’s recommendations are no longer credible to me.
You can avoid this peril of personalization by trying some of these strategies instead:
Capture feedback on the recommendations. For example, Glassdoor could add a thumbs up or down to the email so that users can tell you if the recommendation was good.
Sample some recommendations and manually score them.
Fake it till you make it. It’s hard to build credibility after you’ve lost it, so don’t rush out a fully automated recommendation system. If you don’t have enough data to train the model to a high level of accuracy, then operate it manually or semi-manually first.
Example: Glassdoor Recommendation
When personalization is done right, it can lead to increased sales, conversion rates, and engagement. Done wrong, however, it can feel intrusive or untrustworthy. You can find the balance between intruding on user privacy and providing user value by considering user expectations and user intent.
🗞️ Subscribe to my newsletter
I write about once a week about topics like product management, design thinking, becoming a better leader, and personal branding. By subscribing to my newsletter, you’ll get these insights emailed right to your inbox every time I post.
📹 Reflect with me on Friday Flashes
Friday Flashes is a series of quick one-minute videos that introduces a topic of reflection for the week. Each prompt is product-focused and ties back to one of the ten disciplines in the Product Decagon.
📖 Get your copy of The Productive Product Manager
The Productive Product Manager is a guided journal that combines the weekly reflection with weekly review and other goodies. You can see the inside of the journal in the Friday Flashes announcement video. Don't worry - the journal is not dated, so it's ok if you don't start at Week 1.
