Innovation Through Structure: Mastering the Product Feed with home24

home24 is not your typical online retailer. As a leading home and living eCommerce platform, we offer more than 400K products to more than 60 million users annually in seven geographies and four languages.

This has also forced us to become leaders in product recommendations, running 80+ strategies simultaneously—and increasing our revenue share by threefold in just nine months—our personalization journey to success required a lot of hard work and tough decisions.

The catalysts for change

While home24’s personalization program has grown since its inception five years ago, the past three years introduced a rollercoaster of advantages and disadvantages for our business. From 2020 to early 2022, we experienced a surge in demand as everyone stuck at home during lockdown looked at their old sofa and home accessories and thought it was time to buy new furniture. This naturally led to increased web traffic, conversions, and average order values (AOV).

Everything was great, but then, suddenly, everything changed. Fluctuations in the financial market and fears of an impending recession influenced companies’ valuations, impacting people’s savings. Add in inflation, and consumers were less willing to make purchases, affecting many businesses through shrinking demand and lowered conversion rates.

Internally, things became more complex, too: Our product catalog massively expanded after we acquired Butlers—a strategic partnership to expand our accessory offering—plus the launch of a marketplace. Together, these changes meant we went from managing 60,000 to 300,000+ products in just a few months.

Personalization as our saving grace, with a few tweaks to our infrastructure

On the bright side, personalization offered the company a crucial path forward. It aligned with our vision of offering great value products that people loved through an amazing, delightful shopping experience. But personalization was doubly important now because of its value extraction focus. It could help our marketing spend go further and work harder in an era of rising customer acquisition costs and increased consumer expectations.

Unfortunately, our capabilities were quite limited back then. We’ve always had a recommendations program, using a home-built external infrastructure managed by a team of data scientists, data engineers, product managers, and engineering managers.

Additionally, the infrastructure was limited in terms of speed to market for the tests that we wanted to run since it required a lot of manual work and input from the team. For example, in the best-case scenario, just validating an idea for a new recommendation algorithm would require an investment of three to six months.

Plus, we couldn’t account for the nuances of category type. Recommendations for sofas and curtains should be completely different from one another, as users have different expectations when it comes to the upselling and cross-selling of these products. We were not able to accommodate this with our old infrastructure.

So to make good on the opportunity, we decided we needed a simplified infrastructure that would allow us to deliver effective recommendations in a scalable way using any type of dimension we defined.

From working on just three related projects, we were able to significantly reduce costs and increase our share of revenue by 3X.

Here’s how we did it:

1. Treat product data as an asset and competitive advantage

While fast-growing startups usually focus on launching the product as fast as possible, the fact is that product information integrity always ends up being compromised in the process.

Our team realized the enormity of this problem for the business after having a number of colleagues ask us why a certain product was not showing up as a recommendation. We took a look at the structure of the product feed beyond the market’s standard product attributes and realized that we were working with more than 50,000 combinations of attributes (i.e. “material”) and options (i.e. “leather”) that needed to be carefully selected for each product to deliver the best recommendation.

Many of these attributes were duplicated, irrelevant, inconsistent, or unused or was relevant today but not in the past. For example, while a “sustainability” dimension was not relevant five years ago, it’s now a common attribute for new products. While old products may technically be sustainable, they aren’t marked with this attribute and thus don’t appear as recommendations.

We decided that while cleaning up the product feed would be an incredibly technical, time-consuming, and complicated project, it was also the only way to improve recommendation quality and ensure the data feed’s integrity as the program scaled.

2. Centralize category architecture

Category architecture is the structure of your website, a complex guide of your business reflected in its taxonomy. Like product data, its organization isn’t necessarily prioritized in growing companies, but it’s very important to the customer’s product discovery and how partners ingest your data.

For example: Take “leather sofa.” Ideally, “sofa” is the category, “material” is an attribute, with “leather” selected as an option. This taxonomy will mean it populates when a customer looks at all sofas. However, if “leather sofa” is marked as a category, you’ll isolate all products marked in that category and customers coming to view all sofas will not see leather sofas listed—and this information will also be passed on to your partners.

Simplifying the category architecture for home24’s 300 options took six months and a lot of discussion while trying to answer, “Is this a category or not?” But we didn’t rely on personal opinion for answers. To make the best choices that would not only simplify the lives of our customers, but also enhance our product discovery capabilities, we relied on data we already had, including user search data, marketing campaign traffic data, and filter use data.

3. Mobilize change effectively

Change is painful. I wanted to create the best recommendation experience for my users, and I was told that to do so, I had to completely change my onboarding and rework the categorization and attributes of 300,000+ products. So I started to prepare myself and my team for the long journey ahead.

First, we made sure we had management support, otherwise, we knew we wouldn’t get very far. Then we made sure that we defined a clear working group that was capable of executing the approach for this kind of project. Third, we agreed on the strategy timing and small milestones that would improve results over time.

My best tip of advice? Start with a single, small category, and optimize through small tests and improvements. See what kind of results you get and evangelize the good ones that show value and gain traction.

The journey was not easy, but it pays, which is why our work continues

While we’ve been able to increase revenue from recommendations and provide a solid foundation, the work is truly never done, and there’s so much potential for us to continue growing our personalization program. We’ve set our sights on our next milestone: Increasing our use of behavioral data in product recommendations.

We’re onboarding even more usable and actionable data sets—such as a global feed, keyword fields, and more—to test advanced personalization and differentiate our customer experience. For example, we want to test categories mixed with user types. Just like how recommendation logic is different from sofas to curtains, it’s different from new users to loyal customers as well. By layering these recommendation strategies on top of each other, we can deliver even more meaningful personalization to our customers. We also are figuring out how to account for low incoming data on new products. In comparison to older products with time-tested data, we don’t necessarily know the situations where it will be the best product to display.

Though solving these problems won’t be easy, we know they will pay off with meticulous work, just like our previous bets.