How Netflix’s Recommendation Engine Drives Success

Netflix's recommendation engine decides what almost every member sees when they open the app. It predicts what each person is likely to watch and arranges the whole homepage around that prediction, which keeps people watching and keeps them subscribed.

This article explains how it works, where the famous "80%" figure comes from, why it matters to the business, and what a much smaller product can take from it.

Why recommendations matter so much to Netflix

Netflix has thousands of titles and a member who will give up browsing in a minute or two. If nothing appealing appears quickly, they close the app. If that happens often enough, they cancel.

Recommendations are how Netflix solves that. Two Netflix executives, Carlos Gomez-Uribe and Neil Hunt, set out the reasoning in a paper called The Netflix Recommender System: Algorithms, Business Value, and Innovation, published in ACM Transactions on Management Information Systems. Two claims from it are widely quoted:

• Recommendations influence about 80% of hours streamed on Netflix. The rest comes from search.
• The authors estimated that personalisation and recommendations together save the company more than US$1 billion a year by reducing cancellations.

That paper is the source of the "80% of what people watch comes from recommendations" statistic.

How the Netflix recommendation system works

What it uses

According to Netflix's own explanation of its system, recommendations draw on:

• Your activity: what you have watched, how long you watched it, and how you rated it
• Similar members: what people with tastes like yours have watched
• Information about titles: genre, category, cast, release year
• Context: the time of day, the device you are using, and how long you tend to watch

Netflix says it does not use demographic details such as age or gender.

The methods behind it

• Collaborative filtering. If two members have watched many of the same things, each is shown what the other watched and they have not.
• Content-based filtering. Titles are matched by their attributes. If you watch a lot of dark crime dramas, you are shown more titles with those characteristics.
• Ranking. The system does not only choose titles. It decides the order, because what appears first is what gets played.
• Constant testing. Changes are tested on groups of members and kept only if they improve viewing and retention.

The homepage is the recommendation

Everything on the Netflix homepage is the result of this system:

1. Which rows appear, such as "Continue Watching", "Because You Watched…", "Trending Now" and the "Top 10".
2. Which titles are in each row.
3. The order of titles within a row, and the order of the rows.
4. The artwork shown for each title.

Two members opening Netflix at the same moment see different pages.

Personalised artwork

Netflix has written publicly about choosing different images for the same title depending on the member. Someone who watches many comedies may see an image featuring a comic actor. Someone who watches romance may see the same title's romantic leads. The image is often what decides whether someone presses play, so it is tested as carefully as the titles themselves.

From stars to thumbs

Netflix replaced its five-star ratings with thumbs up and thumbs down in 2017, and later added a "two thumbs up" option. Simpler ratings are easier to give, so more people give them, which gives the system more to work with.

The Netflix Prize

From 2006 to 2009 Netflix ran a public competition offering US$1 million to any team that could improve its rating predictions by 10%. A team called BellKor's Pragmatic Chaos won. The competition brought wide attention to recommendation systems. Netflix later said it did not put the full winning solution into production, partly because the business had moved from posting DVDs to streaming, where what people actually watch is a better signal than the ratings they give.

How the recommendation engine drives growth

• Less time choosing. A short, relevant selection is easier to act on than thousands of titles.
• One thing leads to the next. When an episode or film ends, something likely to appeal is already there.
• More value from the catalogue. Recommendations bring less-known titles to the people most likely to enjoy them, so more of what Netflix pays for gets watched.
• Lower churn. Members who keep finding things to watch keep paying. For a subscription business, that is the number that matters most.

Notice what kind of growth this is. Netflix is not using recommendations to find new customers. It is using them to keep the ones it has. Retention is the growth strategy.

What other products can learn

You do not need Netflix's data or engineers to apply the principles.

1. Help people choose

Too much choice makes people leave. Show a new user the few things most likely to be right for them, whether that is products, features, templates or articles.

2. Start with simple rules

Before any machine learning, simple rules go a long way: "people who used this also used that", "popular with teams like yours", "pick up where you left off". Most products never need more.

3. Use what people do, not what they say

Netflix found that viewing behaviour was a better guide than ratings. In any product, the actions users take tell you more than the preferences they report.

4. Test the small things

Netflix tests artwork. You can test the headline on a page, the first screen of onboarding, the order of options, or the subject line of an email. Small changes to what people see first often matter most.

5. Use context

The time of day, the device and where someone is in their journey all change what is useful to show. A new user and a long-standing one should not see the same thing.

6. Measure retention, not just activity

Netflix judges its system by whether members stay. Judge your own changes the same way: did more people come back?

For how to set up the tracking this depends on, see Product Analytics Tools for Startups. For messaging based on what users do, see Marketing Automation for Startups.

Common questions

How does the Netflix recommendation algorithm work?

It predicts what each member is likely to watch, using what they have watched and rated, what members with similar tastes watch, information about each title such as genre and cast, and context such as time of day and device. It then decides which rows to show on the homepage, which titles go in each row, in what order, and which artwork to display.

Where does the claim that 80% of Netflix viewing comes from recommendations come from?

From a paper by two Netflix executives, Carlos Gomez-Uribe and Neil Hunt, titled The Netflix Recommender System: Algorithms, Business Value, and Innovation, published in ACM Transactions on Management Information Systems. They wrote that recommendations influence about 80% of hours streamed, with the remainder coming from search.

Does Netflix use age or gender to recommend titles?

Netflix says it does not use demographic information such as age or gender in its recommendations. It relies on viewing behaviour, similar members' tastes, title information and context.

What was the Netflix Prize?

A public competition Netflix ran from 2006 to 2009, offering US$1 million to any team that could improve the accuracy of its rating predictions by 10%. A team called BellKor's Pragmatic Chaos won. It drew wide attention to recommendation systems, although Netflix did not put the full winning solution into production.

Why does Netflix show different artwork to different people?

Because the image is often what decides whether someone plays a title. Netflix tests several images for each title and chooses the one a particular member is most likely to respond to, based on what they have watched before.

Conclusion

Netflix's recommendation engine works because it serves one goal: helping each member find something to watch quickly enough that they stay. The methods are sophisticated, but the principle is simple and applies to any product. Understand what each user is likely to want, show them that first, and measure whether they come back.

Gufy helps SaaS and mobile-app startups put that principle to work with product analytics and lifecycle messaging. Book a growth call.