To be the first meal kit service to provide a completely personalised selection of recipes to each customer

Planning healthy, fresh, and easy meals should be a delight, not a chore. That’s why Marley Spoon gives people the choice of more than 20 recipes from a changing weekly menu, and delivers all of the ingredients and recipes they need directly to their door.

Marley Spoon is a multinational meal kit company that has delivered more than 27 million meals to customers in three continents and eight countries around the world over the past five years.

When the company wanted to leverage machine learning to expand beyond its existing 20-recipe menu with the help of personalized taste preferences, I led research, UX strategy, and design for Marley Spoon’s “Taste Profile” project—the first step in unlocking the perfect menu.

The Challenge

When Marley Spoon launched in 2014, only six recipes were offered in each weekly menu. Of these six recipes, two or three (based on their subscription size) would be pre-selected to their order for delivery. Other than customers’ vegetarian preferences set in their settings, these recipes would be selected based on internally set parameters.

In order for Marley Spoon to leave this model and start to utilise machine learning for a truly personalized experience, there were several steps to the vision.

Constraints

With scarce data and limited time, we decided to work on minimum viable product, and build to learn.

Phase 1

Phase 2

In 2018, the company expanded its menu to give customers their choice of twenty unique recipes each week.

While the 20-recipe menu was well-received by customers, Marley Spoon knew that most customers don’t truly have 20 choices due to their own personal preferences and needs. But, expanding beyond the already-large menu adds to an already intensive 12-week process of designing and testing recipes for each weekly menu.

The company was already heavily data-informed, using customer feedback to iterate on original recipes and overall menus to meet a wide range of dietary preferences and needs. That’s why Marley Spoon saw the opportunity to bring machine learning into its operations. It was an intuitive next step to put this data into action first to recommend the recipes that each customer would be most likely to enjoy, and then ultimately, generate dynamic, personalized weekly menus for each customer.

Our product vision set our goal: the second stage of the broader project would involve building a better understanding of our newest customer’s taste preferences, and using this to recommend recipes from our existing weekly menu that best matched their preferences.

Then, we would build on this initial understanding of their preferences based on what recipes they ultimately chose to cook, and the ratings and reviews they provided after cooking each meal, to keep providing better recommendations.

In the long term, we’d be able to predict which recipes would be most popular, and how we could build the perfect menu for everyone.

That meant that the first and most pressing challenge was: How might we measure taste preferences?

Research & discovery

The project team involved stakeholders from every team: product, culinary, development, logistics, marketing, and design.

Collaboratively with the Product Manager, I planned and carried out a research phase to dig into this challenge, including:

• User interviews to learn how people describe their diet and how they choose what to cook and eat for everyday meals.

• Internal interviews and workshops with the culinary team to learn how they think and talk about taste and flavours.

• Ratings and comments gathered from our customer service team and what patterns they held.

• Desk research into existing customer feedback through recipe reviews and support requests, as well as scientific papers around taste preferences, language, and dietary routines.