What We Learned Developing Sunny, a Gift Bot Modeled After Pandora
Gifting is an Art. It requires a unique combination of memory, empathy, intuition, common sense, and perseverance. The act itself is innately human, with its origins deeply engrained within our DNA.
With the advent of AI, we’ve seen vast improvements in the quality of recommendation algorithms. Spotify’s Discover Weekly, Amazon’s Your Recommendations, and basically every page you visit on Netflix are driven by personalized recommendation engines. So, theoretically, this same technology should be able to improve on gifting, right?
Well, as I have learned, gifting is a different animal. Take, for example, UncommonGoods, an online gift retailer, where I lead Product Management. We see users come to our site buying gifts for many different people- their 70 year old mother, eight year old niece, their husband, etc. It’s incredibly hard to “personalize” for this gift giver, because their lifetime click history combines gift searches for many different users, and on their next visit, it will be challenging to guess whether they are shopping for their niece, their husband, or someone else entirely.
This is the problem we set out to solve when we conceptualized Sunny, a gift-finding bot powered by machine learning.
Looking across different recommendation systems, we saw a parallel between gifting and the music streaming service Pandora. Shoppers gift for multiple people, and music lovers listen to multiple genres. On Pandora, users start “stations” by seeding it with a song or artist. We wanted our users to be able to create “stations” that would give them gift recommendations for specific people in their lives.
With a Pandora-esque interface as our initial model, we built clickable prototypes and started to research the viability of a gifting bot. Through user research, we quickly learned many things:
We needed a seed. Like a song that “seeds” a Pandora channel, we needed to allow the user to seed an initial set of products. For this, we leveraged our existing search functionality and offered a handful of popular interests:
Product Consumption moves (a lot) faster than music. On Pandora, songs are served one at a time. We mocked this up, showed to our users, and learned that this model doesn’t work for products. Shoppers have become accustomed to quickly scrolling through category pages and digesting product rows at a time. A serial interface requiring an interaction (click, swipe, etc…) on each product simply does not move fast enough.
Thumb icons don’t translate. The feedback we heard over and over again: “I’m not going to rate a product that I haven’t bought yet.” Users didn’t realize that thumbs up/down was an approval mechanism. They thought it was a binary replacement for the traditional five star product rating system. Ironically, the icon that ended up testing best was, in fact, a single star.
Limiting results could drive feedback. Once we settled on the star icon, we tried many things to encourage users to click stars (to further personalize the product stream). We found that artificially limiting the number of results (thereby forcing them to engage with the star before they could see more) was an effective way of incentivizing the stars and teaching users the value, without having to add a how-to onboarding process. However, it’s a balance. Show too many items, and the user starts to feel like they’re having to do work to get through it — it’s no longer fun. Show too few items, and the user loses faith that they could possibly find what they want — and they move on. We had to find the sweet spot in order to keep them engaged.
Editing interests was a key behavior. In testing, we saw that many users wanted to go back and edit their interests to try and manipulate their results. This ability (as opposed to a complete start over) required a more complex back end structure, but the strong usage rate convinced us it was worth the extra time to launch with this feature.
Down-voting lost value. When we ditched the thumbs, we replaced the thumbs down with an “X” to remove items from the results set. However, we saw that users did not find much value in removing an item from view. In fact, it actually caused them some anxiety (or as the kids call it these days, FOMO). They feared that down voting could unintentionally suppress some great suggestions. Unlike Pandora, where a bad song eats at least two minutes of your life, on a page full of products, you can simply scroll past the ones that are going to make bad gifts. We realized that if down-voting was going to be valuable, we needed to filter out many products, but give users full transparency into the cause-and-effect of their choices. Our future AI vision for this would work something like…The Berenstein Bears:
In “Old Hat, New Hat”, Brother Bear provides extensive feedback to the shopkeeper at the hat store, and eventually, he finds the perfect hat. This will require a sophisticated, nuanced algorithm that we are working towards. To start, we decided to trigger an open text field, so we could understand the type of feedback we could expect to get, and start thinking about acting on it algorithmically.
We needed an animated elephant. OK, so maybe we didn’t actually need one, but we wanted to create a lovable character that would establish Sunny as a being, and embody a more advanced gifting bot. We surveyed our customers on a competing cast of characters including a unicorn and a robot, and the elephant — with its archetypal superior memory — proved the most popular. We animated Sunny with SVG and incorporated these animations into the on-boarding experience:
We released Sunny into the wild this past October, and the initial engagement metrics have given us reason to continue development. In a future post, I’ll share our post-launch learnings and dive deeper into our upcoming features and AI updates. In the meanwhile, give Sunny a try.