Build Recommendation Systems Like the Infrastructure They Are

Amazon’s early “customers who bought this also bought” approach became one of the best-known examples of online recommendations. Behind that feature, the bigger lesson was that product discovery could work as one shared system across an entire store, instead of being built separately for each page.

What Good Recommendations Do for Buyers and Sellers

Good recommendations serve two groups.

For customers, they reduce the work of searching through too many choices.

For authors, publishers, sellers and brands, they provide a way to reach people who may care about what they offer. A product gains little from being listed if the right audience never sees it.

The problem is most in focus for businesses with large catalogs, such as publishers, audiobook platforms and media companies. New releases, first-time creators, specialist products and older titles usually start with less visibility than established bestsellers.

Industry research shows what's at stake.

On the customer side, a 2024 survey found that roughly four-fifths of consumers (~80%) are comfortable with personalized experiences, and many even expect them. Discovering new products was one of the benefits customers said they get from personalized interactions. 

On the business side, a 2021 McKinsey report found that faster-growing companies earned more of their revenue from personalization than slower-growing companies.

The practical lesson: discovery is important enough to manage carefully and test inside each business.

Recommendation Systems Can Solve the Cold-Start Problem

Recommendation systems naturally learn more about popular items because those items are shown more often and produce more information. If no one manages that cycle, popular products keep receiving more attention while new or unusual choices remain difficult to discover.

For a new product, this is often called the cold-start problem: the system has little customer history to learn from. A company must decide how much space to give these items, how quickly new customer responses should change what is shown and how much variety customers should see.

Trying some less familiar choices may look less efficient in a short weekly report. Over time, though, it helps the system learn, keeps the experience from becoming repetitive and gives the catalog room to grow. The goal is to give a relevant new item a reasonable chance to prove that customers want it.

But limited history and missing history can look the same to a recommendation system. A product may already have an interested audience, yet remain difficult to discover because the system can't see that interest. Giving products a fair chance therefore also means checking whether the information used to judge them is complete.

Case Study: How Audible's Missing Data Hid Real Demand

At Audible, I worked on a problem where an audiobook could have an established audience and still appear to have little demand in Amazon search. Even a relevant product can lose visibility when a recommendation or search system has an incomplete view of customer interest.

The same audiobook could have one product record on Audible and another on Amazon. Customers might discover or listen to the audiobook through Audible, but those actions didn't always reach the signals used to rank Amazon listings.

That problem showed up most clearly when an audiobook appeared alongside its Kindle, print and audio CD editions. The other formats could have clear purchase histories while the audiobook’s demand was missing or attributed to a different product record. In some cases, an audiobook bought with an Audible credit appeared to the ranking system as a free item. A customer’s choice had been recorded, but its meaning was lost along the way.

We worked across Audible and Amazon search to connect customer actions to the corresponding audiobook and correct how purchases were recorded. This gave the ranking system a more accurate view of customer interest when deciding which format to show.

Before concluding that customers don't want something, make sure the system can recognize the customers who already do. That applies as much to new releases and niche products as it does to overlooked formats.

AI Shopping Assistants Raise the Stakes for Discovery

AI is changing how people ask for recommendations. Instead of typing a few words and scanning a page of results, a customer can describe a situation, add a budget, explain personal preferences and compare options through a conversation.

Amazon, for example, launched Rufus, a generative AI shopping assistant, to address this trend. The assistant can answer questions about products, compare options and make recommendations based on the conversation. 

A carousel may show 10 choices, while an AI assistant may suggest only one or two, so each recommendation carries more weight. Businesses need to understand what the assistant is trying to achieve and which rules shape its answers. They should also check whether it helped the customer find something useful or simply sent more attention to products that were already popular. 

A Better Scorecard for Recommendation Systems

Clicks are useful, but they can be misleading. People often click what appears first or what they already recognize. Sales can also favor well-known products. A company can improve short-term numbers while customers see less variety, smaller suppliers lose visibility or trust declines.

A simple scorecard should look at four areas:

• Customer experience: Did people find something useful, save time, avoid unwanted returns and come back?
• Opportunity for suppliers: Are new, specialist and older items reaching the people most likely to value them?
• Business results: Did the recommendation create sales or repeat use that probably would not have happened otherwise?
• System reliability: Is the experience current, available, affordable to run and easy to correct when something goes wrong?

No single number can answer all four questions.

Teams also need clear ownership: who decides what the system should optimize, who reviews whether a handful of products get too much attention and who can pause a change that harms customers or suppliers.

Case Study: What Audible Learned From Listening Behavior

One lesson from my work was that a purchase does not always show what a customer truly enjoys. In a basic purchase-based system, every purchase carries the same weight. It can see someone purchased a book, but not whether they stopped reading it after 10 minutes or bought it as a gift.

For audiobooks, we built a similarity dataset using listening behavior. It considered:

• How long someone listened
• How much of the audiobook they completed
• When they listened

A person might prefer a business book during a morning commute but a novel before going to sleep. We also tested signals from other formats. A Kindle or audio CD purchase could suggest interest in the same title on Audible. A person who watched "The Martian" might also be interested in the book. These connections were useful, but they were still indirect.

In our tests, listening-based recommendations worked better than purchase-only and cross-format approaches. The reason: listening showed what people actually spent time with, not just what they bought. The tests showed how recommendation systems improve when they learn from real customer behavior and use those lessons across the business.

Treat Recommendations as Shared Business Infrastructure

Recommendation becomes infrastructure when one dependable system can support search, browsing, AI assistants, customer messages and partner experiences. The system can share customer preferences, catalog information, rules about what may be shown, testing methods and ways to measure results. Each channel can still present recommendations in a way that fits its audience.

This requires clear teamwork on top of better technology:

• Product leaders define the customer and business goals.
• Engineering and data teams build the shared system.
• Business and content teams explain which parts of the catalog need support.
• Privacy, security and legal teams set boundaries for how information may be used.

Everyone should work from the same plain-language definition of success.

The most important shift is how leaders think about the system. A recommendation is not just a guess about what someone will click. It's a decision about which parts of a business receive attention. When that decision shapes many customer journeys, it deserves the ownership, measurement and care expected of other critical business infrastructure.

Learn how you can join our contributor community.