Idea
While exploring AI capabilities, our engineers created an MVP of a chatbot trained to recommend books based on user input. It showed strong potential as a solution.
Book recommenders
Using AI to help bookworms find their next read
10%
Conversion via AI recommender
+6%
Boost in registrations
10×
Increase in recommender usage
Challenge
At Martinus, an online bookstore, we learned that many bookworm customers struggle to decide what to read next.
They often told us:
‘There is so much to choose from, but which books will I actually like?’
My team’s goal was to help undecided readers choose a book and increase conversion.
Solution 1
AI recommender
Testing the MVP
Before making any changes to it, I suggested shipping the MVP and learning from real users first. To clarify and prioritize what we wanted to learn, I led an Assumptions mapping workshop.
Assumptions
Learnings
I gathered feedback through usability testing and an after-use survey. We also ran an A/B test and analyzed the prompts.
- Users who used the chatbot converted slightly more often than those who didn’t.
- Most users found the recommendations helpful and relevant.
- No users raised concerns about AI sustainability.
- Many users asked customer support questions instead of requesting book recommendations.
- Some users complained about the small size of the chat window.
- Some users found the chatbot too intrusive.
- The chatbot’s responses were slow, which frustrated some users.
New concept
Seeing the potential to increase conversions, we decided to address these issues. I designed a dedicated chatbot section that isn’t limited to a small chat window and no longer looks like customer support.
We named the chatbot Sherlock and replaced a feature called Bookworm Helper, in which customer care employees assisted readers with book recommendations.
Making it discoverable
For the chatbot to be useful, users need to find it – and a random ‘sparkles’ icon won’t do. I designed several unintrusive entry points and placed them along the customer journey. This step boosted the number of Sherlock users by almost 900%.
This is where many undecided readers start by choosing their favorite genre.
Many people use pagination when exploring books, making it a fitting place to invite them to try the chatbot.
The same applies to the homepage, where users can enter prompts directly.
Prompt analysis
About 75% of users use Sherlock as intended — to get book recommendations. The rest use it differently, so we adapted accordingly:
Some users coming from homepage mistook Sherlock for search. Although this dropped after adding an animated placeholder, we still trained Sherlock to handle it.
Some users asked about their orders or accounts. Sherlock now guides them to the right pages.
Sherlock currently doesn't have pricing or stock data. We plan to add this later.
Further feedback
I reviewed conversations that received poor ratings and identified patterns that we plan to address in future updates:
- Poor age targeting
When customers ask for books for teenagers, young adults, or seniors, the results often miss the vibe they had in mind. - Asking for similar books
When users ask for books similar to something they like, Sherlock often suggests the exact same book or the same author instead of something new. - Fixating on minor words
Sometimes Sherlock fixates on an unimportant word in the prompt and only provides results related to it.
Impact
10% conversion rate
Among Sherlock users, 10% purchased a book it recommended.
Reduced workload
Customer support no longer handles book recommendation requests (without a layoff).
Tool for clerks
In-store clerks use Sherlock to recommend books to customers.
Solution 2
‘For you’ recommender
Observation
When browsing books, I noticed that many research participants use the ‘Similar books’ carousel on the product detail page. They use it as navigation to explore new books.
Idea
I suggested to create a page dedicated to personalised recommendations (now improved with AI), based on books the customer has previously viewed or purchased.
Testing the MVP
Working with the developer, we explored the simplest way to implement the MVP. I decided to reuse the same component, enhance it, and repeat it several times on the new page. This approach increased loading time, but allowed us to quickly test the concept.
Assumptions
Learnings
I gathered feedback through usability testing and an after-use survey. We also tracked user engagement and conversions.
- Users who used the ‘For you’ section converted slightly more often.
- Most users found the recommendations relevant to the books they were based on.
- Many users requested recommendations based on more than three purchases.
- Many users didn’t log in and therefore didn’t see personalized recommendations.
- Some users found recommendations irrelevant when the books they were based on were gifts.
- Some users said they would appreciate recommendations outside their usual genre.
Improving on the go
Encouraged by the initial results, we rewrote the code to improve loading speed. We then gradually introduced several UX improvements that I designed:
Further feedback
- Already owned books
Many users complained about seeing books they already own in the recommendations. We plan to filter these out. - Repeated books
Users who read within a single genre often saw the same books repeatedly. We plan to address this by improving the algorithm. - Lesser-known books
The algorithm favored bestsellers, which caused lesser-known books to be underrepresented, despite readers actively seeking them. We plan to test ways to balance this.
Impact
4.5% conversion rate
Among ‘For you’ users, 4.5% purchased a book recommended by the feature – well above the site average.
Increase in registrations
The feature provides clear value to logged-in users and boosted registrations, with 6% of visitors signing up within 15 minutes.