From searching on Google, shopping on Amazon, jamming on Spotify and chillin with Netflix, the intrinsic benefits of personalization can be felt by us all as consumers. Often times the best user experiences are described as a seamlessly wonderful and almost invisible experience which fades into the background of using that product. If you’re a customer-centric organization looking to create these magical moments within your product, then incorporating this philosophical approach to customer engagement will do just that. Nearly every product or service can be personalized to some degree and we’re seeing personalization trends on the rise across multiple markets and industries. If you’re not leveraging the power of personalization currently, then your users already probably know that and are just waiting for a competing offering to come along that does. The business impact that personalization can have on a product is too great to ignore, and doing so will leave your company at a severe disadvantage. At the heart of creating these satisfying experiences lives an endless list of personalization techniques, requirements and considerations so that presenting the right content at the right time to the right user can be accomplished. And as a technologist responsible for building these experiences, I’ve seen modern day technology and optimization techniques evolve to a point where it’s possible to build highly optimized, customized and personalized user experiences at a scale that we couldn’t imagine a decade ago.

Now we can’t talk about building a sophisticated system that enables personalization without discussing the primary concern that users have when they know a product is being personalized for them — the privacy of their user data. Companies need to carefully design their personalization experience because as Uncle Ben from Spider-Man famously said, “with great power comes great responsibility”. How your company collects, stores, and utilizes this data should be a top priority and consideration set when designing any personalization capability. In addition, global regulations such as GDPR are forcing companies to harness user trust in a way that is compliant, secure, and safe.

In this post, I’ll outline common steps and considerations when building a personalized product experience which will get you on the pathway to making personalization a core competency for your company. Each section could be a multi-day workshop itself, rather we’ll just scratch the surface of these sections from a practicality standpoint.

Personalized recommendations need to be diverse enough to account for the wide spectrum of a user’s interests.

Defining Personalization

A popular personalization equation is simply articulated as the identification and categorization of a specific user type + the delivery of the most appropriate content = maximized user satisfaction. However, personalization can also be implemented in such a way that goes beyond the content itself. It can even be thought of as a feature. Companies can personalize their product experience by allowing the user to define or refine their experience. Uber allows users to create a ‘home’ and ‘work’ address so when that user open the app up it quickly suggests those locations first as their suggested destination. Netflix allows their users to create different profiles such as a Kid profile which restricts PG-13 and above content to allow a child to easily navigate the product with a more personalized and safe experience. Many B2B products that I’ve worked on allow a lot of customization within their products by giving their users the ability to remove or add particular reports or capabilities within the product. This is commonly achieved in an Admin Permissions section of their product. So when you think about Personalization, try to think more broadly than just the content that is being served up to a particular user. For the purpose of this article, I’ll focus on the use of personalization within B2C products.

To actually achieve the desired outcome of personalization, one needs to understand the complexities and intricacies that make up that overly simplified equation. Let’s dive into more of the specifics that can turn this ambitious dream into a reality.

The Importance of Data

The quality of your personalization engine begins with data and your analysis of that data. Your ability to collect, store, and understand the right data that exist for each user or user cohort needs to be at the center of that product strategy. At a minimum, a dedicated product manager and data analyst should practically live in your virtual data warehouse to fully understand what explicit and implicit user data exists and what data requests are not being captured today. I’ve also witness companies struggle to build a successful personalization product experience because they have paralysis by analysis. Determining what is the right data can sometimes be an art and science — but ultimately always leads back to understanding your users. One of the best techniques I’ve used for understanding a user is by creating a customer journey map. This exercise helps you outline every step that your user takes related to their user problem and their journey to seek a solution to that problem.

Customer journey maps help you understand the physical and emotional steps that a user goes through as they seek solutions to a specific user problem.

Sometimes the data you need sits outside of your team or product — which is why a cross-functional and organization-wide discussion is critical when discussing the topic of data needs. Each product or service differs from the next but a good place to start is by looking at existing behavioral data, demographic info, user settings/preferences, search/purchase history, and content information (such as ratings, transaction history, product specs, product category).

Once your team has established a solid understanding of what core data exists, your team then needs to ensure that data source is accurate, up to date, and reliable. All subsequent steps will be based on this data — and we’ve all probably experienced the negative aftermath when personalization has gone wrong. You purchase an online gift for someone that clearly doesn’t share the same interests as you — but next thing you know, that’s all the company wants you to see. Knowing how to handle false positives and filter out negative signals is just as important as the positive indicators.

Probably the most advanced form of personalization is being utilized at Netflix, where they themselves think of building 100 million products, not just one product. Every tiny detail is customized and personalized. And what is being customized is based upon a user’s behavioral patterns and their likelihood of watching a show or movie. We’ll get into the details of this a little later, but you need to think about the data you can collect which will inform your experiments in subsequent phases.

Carefully Architected Design System

It should be obvious by now, but you will need the ability to house huge amounts of data that run through sophisticated machine learning algorithms that continually get more complex as your algorithms evolve and mature. This requirement calls for a carefully architected software design system that enables flexibility and agility since the business, industry, and environment will constantly be in a state of flux. If additional data points or sources need to be added, modified, or removed you’ll need a system that can do that easily. A thoughtful approach to designing this system is needed upfront to mitigate the likelihood that in a year you’ll hear the dreaded ‘refractor needed’ request coming from your engineering lead. At the same time, it’s also not realistic to think that whatever is built today will stand the test of time — so there is a careful balance there. Try thinking about the number of experiments you’ll want to run, how you’ll run them, what pieces of your product could be customized or personalized, and where your personalized product could be in a year or two from now if you succeed. Designing this system correctly should be a collaborative process with many eyes on it.

User Identity and Categorization

If we go back to the simplified personalization equation outlined above you will notice the first component is to identify and categorize the user type. Before you begin to actually personalize a users experience you need to know with high confidence what content will resonate most for that user. This is the step which joins all of the data you have on that particular user with the predictive algorithms you’ll create. Tagging a user correctly is very important because as mentioned earlier, we probably have all experienced a miscategorization before. This leads to the deterioration of a user’s trust in the brand or product. If you completed your customer journey map exercise then you also probably have a good understanding of trigger-based actions or events that can be included in this categorization of your user. For example, a customer moving to a new home means that despite their past search history or behavior they’ll most likely be in the market soon for a new cable/internet provider.

Create Ranking Algorithm Hypotheses

I can’t emphasize enough the importance of creating your algorithm hypotheses with a cross-functional team. Companies must collaborate with team members from user insights, engineering, product management, marketing, operations, etc to craft potential ranking approaches. Data insights will coalesce with your foundational understanding of a user’s journey to be eventually translated effectively into an algorithmic hypothesis which will be later tested and optimized. Let’s look at an example of Netflix to better understand how these hypotheses play out in their product. They articulate much of their personalization testing strategy on their Medium Netflix Tech Blog, so if you’re interested in knowing more check that out.

An example of a personalized Netflix homepage.

When you look at the Netflix homepage it’s hard to imagine that they have nearly 100+ million variations of this page, but it’s true. Here is how they think of how and what can be personalized on this one page:

Which row categories to display (top picks, trending now, because you watched, new releases, specific genre categories, etc)
Of the row categories that Netflix decides to show you as a user, they then need to determine the order and position of those rows. Should Top Picks be at the top or below Trending Now?
Within that selected row which movies or shows are shown to a user?
What about the position of those movies/shows within a given row? Should this movie be position 1 or 5?
Once the row, sort order, position, and content has been determined — the next question is how should they display this piece of content to a specific user. This is where it gets really remarkable. For example, maybe you as a user are drawn more towards scenic images vs. the actors themselves. Or maybe you have a favorite actor who is in the show or movie that you’re looking at. All of that is taken into consideration to determine which thumbnail is displayed for any given selected piece of Netflix content. Take a look at the images below to see examples.

Netflix personalizes hundreds of movie thumbnails for a given user.

Based upon a user’s historical actor preferences (Uma Thurman or John Travolta) the Netflix image algo will determine the best way to display movie/show titles to a user.

Testing Ranking Algo Performance

At this stage, you feel confident that you’re collecting the right data for a user and you have several hypotheses about what will drive the best possible user experience. For this stage, a calculated approach of thoughtful data and performance analysis needs to drive your decision making. One of the biggest obstacles preventing companies from succeeding with personalization at scale is that they don’t construct the right system that allows for testing a broad set of ideas quickly. So before you begin testing you need to know what will define success or failure. Start by agreeing on a set of success metrics prior to any testing. Here’s how Spotify initially structured their success metrics:

Reach: Discover Weekly Weekly Active Users / Spotify Weekly Active Users = what % of weekly active users are using Discover Weekly playlist?
Depth: Discover Weekly Time Spent / Spotify Weekly Active Users
Retention: Discover Weekly week-over-week retention

Once you’ve agreed on the success metrics that you’re shooting for, there are countless techniques one can use to finalize the algorithms that drive the best-personalized end-product. Here are a few to get you started:

Collaborative Filtering: This is the most basic form of a matching algorithm whereby you make a prediction about the interests of a user by matching that particular users’ past behavior with information and interests from many other similar users. The key here is that similar users share the same interests so they will commonly like the same products or content.
Natural Language Processing (NLP) Filtering: Through incorporating other data or sentiment analysis you can begin to decompose the interests of a particular user to make intelligent future recommendations. For example, at ClassPass we collected ratings and reviews of each fitness class completed by our users. Within those reviews of a fitness class lived a goldmine of data that could be utilized to actually correctly categorize the class and/or the users' interests. Users would mention they loved the Electric-dance music in the class. We can then use that information about that user to match them with other classes we know have Electric-dance music.
A/B Testing: a most traditional approach for comparing the results from one test (A) against another test (B).
Interleaving: This is a powerful but much more complicated technique that has enabled companies like Netflix to accelerate their ranking algorithm innovation. There are many detailed how-to guides that I’d recommended reading if you’re interested in accelerating your experimentation process for testing new algorithms. Essentially it’s a process for fast pruning many tests to identify the most promising ranking algorithm from a large initial set of ideas.

Establish Ethical Principles

As an organization, you cannot skip this critical stage which will set the stage and also provide guidance towards future ways of building and optimizing your products. At the heart of what you’re trying to accomplish is to provide products or services that solve user problems. Without your users, you don’t exist. So your user’s security of data and privacy of that data should be at the center of everything you do — and protecting them and their concerns should be a top priority. This will bring up incredibly difficult and challenging topics that might, in fact, directly compete against your revenue models. This is exactly why you should go through this exercise and outline your company’s position on particular security and privacy topics.

Conclusion

I have just scratched the surface as it relates to this incredibly important and impactful topic of building personalized customer experiences within your product or service. My hope was to provide you with a basic framework and baseline understanding so you can seek out advanced algorithms for building a recommendation and personalization system. If you have any questions or comments don’t hesitate to leave a comment below or drop me an email: kevin@kevinbeatty.com. If you like this post, make sure you follow me on Twitter for more product related insights and thoughts.

How to build Personalization as a core product competency for your company.