Prioritizing user trust to boost engagement by 50%

The original web app was built for small wineries who uploaded their full inventory to PairAnything. From that list, our algorithm would recommend two wines at a time. However, Target was our first major retail customer. Instead of uploading their entire catalog, they wanted to promote just two featured wines. If we used the old setup, those same two wines would show up no matter what a shopper searched for. Our credibility was at stake, and we were at risk of breaking user trust or losing them altogether.

If shoppers didn’t believe our recommendations were genuine, it would make Target look less knowledgeable as a wine brand, and jeopardize our partnership. This called for a redesign, where we explored how many and what wines to recommend to users.

More concretely, I wanted to know:

• How many pairing recommendations do we show? Do we allow users to “refresh” or “Show More” pairings?
• Do we prioritize Promoted wines?
• Do we show wines of each type (white, red, etc.)?

Through iterative rounds of prototyping and feedback from the CEO, sommeliers, and users, I learned:

• 4 is the perfect number of recommendations.
• We should avoid reinventing the wheel and leverage Target.com as much as possible.
• Users preferred genuine pairings, without any special rules regarding wine type.

With the old design, I heard a common complaint from users:

I hypothesized that showing more than two recommendations, or letting users “refresh” for a new recommendation would make the experience feel organic and give room to show more diverse wines. I wanted to know:

To explore this question, I prototyped three different versions of the recommendation page:

• one with 10 recommendations;
• one that lets users click ‘Show More’ to see a total of 4 recommendations;
• one with 3 recommendations, and an option to ‘Refresh’ to get a new set.

Usability testing revealed that:

• Four recommendations was ideal. 10 was too much.
• Users didn’t want to click an extra button (”Show More”) to see all pairing recommendations, especially if it was only four results total.
• The “refresh” interaction made the recommendations feel random instead of intentional, which could damage our credibility.

Following user feedback, I decided to show four pairing results, and omit a “refresh” or “show more” interaction. Additionally, I talked with the CEO and advocated to recommend all Target wines to users, regardless of promotion status, so that Pinot Noir and Riesling aren’t the only wines that users see (more on that below).

In the old design, we would only recommend Promoted wines, causing users to view our service as an ad. This raised a key question:

My goal was to build a recommendation platform that users can trust, while keeping the technical scope to a minimum for our solo-developer. What if instead of limiting recommendations to Promoted wines, we leveraged Target's existing online store's search function to recommend all wines sold at Target? This way, users are driven to Target.com for every recommendation, and no extra technical integration with Target is needed, keeping the scope to a minimum for both PairAnything and Target.

Now, all wines sold at Target are considered for recommendations, regardless of Promotion status. Every recommendation comes with a “Buy at Target” button that takes users to a Target.com search for that wine. Promoted wines are showcased under a “Get it at Target” section with direct links to their product page on Target.com. Additionally, promoted wines get a dedicated detail page within PairAnything.

With the redesign, we deliver value to both Target and their shoppers:

We applied the same pattern of leveraging the Target.com search function for wine-to-food searches. When users search for a wine and receive food pairing recommendations (ex. searching "Cabernet Sauvignon" returns "Beef Brisket"), the main CTA button labeled, "Shop at Target" drives users to the Target.com search results for the dish name, which displays purchasable ingredients for that dish.

In conversations with Target, they raised a key concern:

This lead us to explore:

To test this, I prototyped a version where every recommendation set included four wine recommendations, categorized by type (red, white/rosé/sparkling), and at least one wine from each type.

Early design where at least one wine of each type is recommended to users.

I usability tested this prototype and learned that both sommeliers and the wine novices found this approach confusing. They questioned the integrity of the recommendations, and found it harder to pick out a wine.

Rather than imposing rules on wine types, I advocated to keep recommendations genuine. To address Target’s concern, “What if the shopper doesn’t like red wine?”, I introduced Like/Dislike buttons. Shoppers can now personalize their pairing experience by telling us why they dislike a recommendation, and we’d remember their preferences for future recommendations.

This approach allowed us to offer genuine pairing recommendations while positioning Target as a reputable wine retailer. Additionally, the ‘like/dislike’ feature lets us tailor recommendations to shopper preferences, paving the way for AI personalization, a long-term goal for PairAnything.

Beyond improving how recommendations were made, I addressed core usability issues with the pairing experience identified in previous user research interviews. In the old design, users had to choose ‘Wine’ or ‘Food’ first before typing their search, causing failed wine searches under ‘Food’ (ex. searching ‘Chardonnay’ under ‘Food’), leading users to exit the app without seeing any pairings.

In the redesign, I combined food and wine searches into one search bar with autofill suggestions from both categories. Now, users can simply type their query, and receive the correct pairing. I also made the search bar sticky at the top on all pages, boosting the engagement rate from 5% to 56%.

The old design required users to tap ‘Food’ or ‘Wine’ first, then type in their search; causing users to leave due to failed search attempts. In the redesign, users can simply type in their search and see autofill results for foods, wines, and promoted Target wines.

When testing the old design, we heard that users wanted:

• Confirmation that PairAnything understood their search.
• More context about the recommended wines and why they pair.

So I redesigned the recommendation screen to:

• Display photos of the wine and food, confirming we understood their search.
• Include a short description about why the wine and food pairs.
• Visually present how the wine tastes with images of key flavors and scales.

The final design displays visuals of the food and wine pairing (1), explains why it pairs (2), and describes how the wine tastes (3).