Personalization

011 min

The problem personalization solves

A Netflix member sits down in the evening to find something to watch. The catalog holds thousands of films and series, far more than one screen can show. Netflix researchers Carlos Gomez-Uribe and Neil Hunt reported in 2015 that a typical member gives up after 60 to 90 seconds of looking, by which point they have seen 10 to 20 titles. If nothing has caught their interest by then, the chance that they cancel the service goes up.

So the home page cannot be the same for everyone. A member who watches crime documentaries and a member who watches cartoons with their children need different titles on those first two screens. Netflix builds each member's home page for that member: which rows appear, in what order, and which titles sit in each row.

That is personalization at work. The rest of the story is about how a system makes those choices, when the choices help, and when they get in the way.

023 min

How personalization works

Netflix's home page, like every personalized screen, rests on three parts: signals about the person, a rule or model that turns the signals into a choice, and a place on the screen where the choice appears.

• Signals. Some are explicit: the genres a new member picks at sign-up, a thumbs-up, a saved home city. Most are implicit: what the person watched, bought, clicked, searched for or skipped. Context counts too, such as device, time of day and rough location.
• Rules or models. The simplest systems use rules a person writes, such as "show visitors from this city the opening hours of the nearest store". Others put people into groups first, by role or by segment, and give each group its own version. A learning app might show teachers a class overview and students their next lesson. The most complex systems use machine learning to predict, item by item, what this one person is likely to choose.
• Placement. The output can change content (which titles, products or articles), order (what comes first), layout (which blocks appear) or message (the subject line of an email, the offer in an ad).

On Netflix, the signals are mostly what each member has watched and rated. Several ranking algorithms score the catalog for that member. A page-generation algorithm then picks the rows, typically about 40 of them with up to 75 titles each. A "Because You Watched" row is built around one title the member already finished.

Then the loop closes. What the member plays from tonight's page becomes a signal for tomorrow's page. This is why personalization improves with use. It is also why it can drift: the system learns mostly from items it already chose to show.

Item-to-item matching at Amazon

Much of this rests on collaborative filtering: suggesting items to one person based on the choices of many other people. The older form looks for other customers whose purchases overlap with yours, then suggests what they bought. That search gets slow when a store has tens of millions of customers.

Greg Linden, Brent Smith and Jeremy York described Amazon's answer in 2003. Their item-to-item collaborative filtering starts from each product you bought or rated and looks up products that other shoppers often buy along with it. Amazon builds that table of related products ahead of time, so little work is left for the moment the page loads.

The result is a store that looks different for each visitor. A software engineer and a new mother could open the same store and see different shelves: programming books for one, baby toys for the other.

Choosing the picture and the reason

Personalization also changes how a single title is presented, not only which titles appear. Netflix calls this evidence selection. One algorithm decides whether a recommendation should mention an award, or say that the title is like something the member watched last week. It also chooses which of several images of the title to show. Two members can see the same film with a different picture and a different reason to press play.

031 min

Where personalization comes from

Personalization has no single inventor. It is practitioner convention: the word spread through web design and marketing in the late 1990s, once sites could remember visitors through accounts and cookies.

The clearest early line between the terms came from usability researcher Jakob Nielsen in a 1998 column. Customization, in his use, was under the user's direct control, such as choosing which stock prices a portal page tracks. Personalization was the version run by the computer, which builds some model of what the user needs and serves pages to match it. Nielsen thought personalization was over-rated. In his view, it was often an excuse for not building a site people could simply navigate.

The following years gave examples for both sides of that doubt. Amazon's recommendations, which learned from purchases with no setup, became the standard case of personalization that works. Microsoft Office 2000 shipped menus that changed for each user, and became a standard case of the opposite. Both come back later in this article.

042 min

Personalization vs. customization and segmentation

Nielsen's split still matters, because three terms are often used as if they meant the same thing. They differ on one axis: who decides what each person sees, and for how many people at once.

Customization asks the user to do the work, and most people do not. Nielsen called this the paradox of the active user: people want to start using a product, not spend time setting it up. Segmentation is often the first step toward personalization. A rule for a group is easier to build and explain than a model for each person, and many products that say they personalize are really segmenting.

A recommender system is narrower than personalization. It is one tool personalization uses, for choosing items. Personalization also covers order, layout, messages and defaults.

What users ask for is not what works

One common reading says personalization should give people more of the choice they ask for. Netflix found the reverse. Gomez-Uribe and Hunt wrote that members ask for as much choice as possible, plus full search and browsing tools. What actually worked, in Netflix's own tests, was a few strong options presented simply. This matches the research on choice overload: more options can make a decision harder. Here, the job of personalization is to remove options, not add them.

052 min

How personalization shows up in real products

The Netflix home page is the clearest published case of personalization working at scale, and the same pattern shows up across the products people use every day.

The other 5% of the time, the person wants somewhere else, such as a city they are about to visit. So the site still needs a plain way to choose another place. Personalization sets the starting point. It does not remove the choice.

Where the need moves: one account, several people

The Netflix paper names the harder case. Many accounts are shared by more than one person in a household, and the same person wants different things on different evenings. A model trained on the whole account cannot tell whether tonight's viewer is the parent or the child. Netflix responds by showing a wide mix of rows, so a viewer can skip the rows meant for someone else and find one meant for them. Separate profiles inside one account help in the same way: they tell the system who is watching.

062 min

When personalization backfires

The household problem points to a general rule. Personalization fails when the system's guess is wrong and the person cannot easily get around it.

Menus that move: Office 2000

Microsoft Office 2000 opened each menu in a short form that held only the commands the user had used recently. The full menu appeared after a pause, or after a click on an arrow at the bottom. Commands moved in and out of the short menu as habits changed, so a command could be missing from the place where it was yesterday. Office 2007 replaced these menus with the ribbon.

Leah Findlater and Joanna McGrenere at the University of British Columbia tested this kind of design in 2004. 27 participants used three versions of a menu. In the static version, the top four items never changed. In the adaptive version, the system filled the top with items the participant used often or recently. In the adaptable version, the participant chose the top items. The static menu was significantly faster than the adaptive one. Most participants preferred the adaptable menu, where they were in control.

071 min

Does personalization create a filter bubble?

Over-personalization raises the larger worry about news and social feeds. The fear is that a feed ranked by personal data shows people only views they already hold. This is often called a filter bubble, or an echo chamber.

A 2015 study in Science by Eytan Bakshy, Solomon Messing and Lada Adamic measured this on Facebook, among US users who listed a political leaning. They followed hard news stories through three steps. Of the stories that people's friends shared, 29.5% crossed political lines. Of what the News Feed showed them, 28.5% did. Of what they clicked on, 24.9% did.

The News Feed ranking did lean slightly toward each user's own side. But the friends people chose, and the links they chose to click, narrowed what they saw more than the ranking did. All three authors worked at Facebook, the company whose system they measured, so the study is not independent. It still corrects one misreading: the bubble is not only the algorithm's work. A design that adds variety to a ranked feed helps, but it cannot fully make up for a narrow circle of friends.

081 min

Personalization and accessibility

Moving items around has a second cost, for people who depend on a page staying the same. WCAG 2.2 Success Criterion 3.2.3, Consistent Navigation (level AA), requires that navigation repeated across pages keeps the same relative order, unless the user starts the change.

A navigation bar that reorders itself to match each person's habits, without being asked, works against this criterion. The people it affects most:

• Screen-reader users who have learned the order of the links.
• People with memory or cognitive disabilities, who rely on finding things where they were.
• People with low vision who magnify the screen and see only part of it at a time.

The criterion still leaves room for personalization the user controls. Changing the order through settings or an assistive tool is allowed. This is the accessibility argument for the adaptable menu in the Findlater study and against the adaptive one. It is a specific case of the wider heuristic of consistency and standards.

091 min

Personalization and privacy law

Privacy law sets outer limits on the data personalization can use. The EU's General Data Protection Regulation (GDPR) and California's Consumer Privacy Act (CCPA) give people rights over the personal data a business holds, including the right to find out what it holds. A design that hides which data shapes the page makes those rights harder to use.

The Digital Services Act opt-out

The EU's Digital Services Act (DSA) of 2022 goes further for the biggest services. Article 38 requires very large online platforms and very large search engines to offer, for each recommender system, at least one option that is not based on profiling. Profiling here means the automated use of personal data to judge or predict things about a person.

For a product of that size, a view without personalization is a legal requirement, not an extra. For smaller products, it is still a useful test. If the product does not work without personalization, the default design is too weak.

102 min

Applying personalization

The cases above add up to a short list of design moves.

1. 1

   Design the default first.

   New users, and anyone the system knows little about, see the default. Nielsen argued in 1998 that even a site built on personalization needs a good default design for first-time visitors. Because of the default effect, most people will also keep whatever that default is.

2. 2

   Start with rules and segments.

   Personalize one or two places where a rule is easy to state, such as location or role, before building a model for each person.

3. 3

   Ask, then watch.

   A short question at sign-up, such as picking a few genres, gives the system explicit signals while it has no history.

4. 4

   Copy, do not move.

   Put suggested items in a separate, clearly labelled area. Keep menus and navigation in a fixed order.

5. 5

   Say why.

   A label such as "Because you watched…" tells people which data drove the choice, so they can judge it.

6. 6

   Give control back.

   Let people correct a suggestion ("not interested"), reset their history, or switch personalization off.

7. 7

   Keep some variety.

   Reserve a few slots for popular or new items the model would not pick, so the page does not narrow over time.

8. 8

   Join the data.

   Signals split across the website, the app and email give each channel a different picture of the person. One shared customer record keeps them consistent.

9. 9

   Test against the default.

   Compare the personalized version with the plain one in an A/B test. Judge it on a lasting outcome such as retention, as Netflix does, not only on clicks.

10. 10

    Keep a person reviewing it.

    Someone should check what the system shows in sensitive categories, because a model does not notice when a correct guess is unwelcome.

Picture a team at a grocery delivery app deciding where to start. A safe first move is a "Buy again" row above the category menu, filled from each customer's past orders. The category menu stays in the same order for everyone, so a customer who wants something new still finds it where it always was.

111 min

How to spot personalization going wrong

Personalization that has gone wrong leaves traces you can check in recordings, tickets and analytics.

• Session recordings or support tickets show people hunting for a command or page that used to sit in a fixed place.
• Many people switch personalization off, reset their history, or press "not interested".
• Clicks on the personalized area rise while retention or purchases stay flat.
• Recommendations repeat items the person already bought or finished.
• Feedback uses words like "creepy" or "how did it know".
• One profile's suggestions mix two very different tastes, a sign of a shared account.

?4 questions

Questions people ask

§10 sources

Sources on personalization