We've Entered the Era of Hyper-Personalization

Walking through the looking glass

“Dream is the personalized myth, myth the depersonalized dream”

- Joseph Campbell

We’re entering a new digital era where the world around us will be hyper-personalized. What is hyper-personalization? Hyper-personalization is the creation and delivery of unique content tailored to a specific individual. It goes beyond serving content that an individual might like to specifically creating content that appeals to that individual. Different words and images for each individual, that's hyper-personalization. If you don't think we are there yet, surprise! AI has cleared the current bottleneck to a hyper-personalized world.

What does that look like in practice? Carvana just released a hyper-personalized campaign in which over 1.3M unique thank you videos to were sent to each customer who purchased a car with them. This is what the results look like (note the videos may not work directly in email due to a Substack issue, here is a link to view the videos):

Here’s a different one:

If you enjoyed those, you can find more videos here. Carvana has the first large scale campaign using hyper-personalization that combined information about individuals with content generated to a single individual. In Carvana’s own words:

Ultimately, you get what you see on-screen: A video tailored specifically to Holly, created by a system capable of generating over a million other unique videos for each and every customer we’ve had the pleasure of selling a car to. 

Why do we want to hyper-personalize? Well, personalization has been shown to greatly improve engagement, response rates, and general company performance. People personalize content and messages all the time, they just don’t realize it. If you’re headed to a theme park for a vacation, you’re going to explain it differently to your spouse vs your kids vs your friends vs your employer. Let’s look at how personalization has progressed over time.

A Brief History of Personalization

Let’s say you were having a conversation with someone and wanted to show them an image of “Michael Jordan”, how would you determine which one to show? Would you pick the basketball player?

The computer science professor?

Or perhaps the actor?

Who you picked would depend on the context of the conversation and who the individual is. If they are thespian, chances are you’re more likely to show them Michael B. Jordan the actor. If you’re having a conversation about advancements in AI, you’re likely to show Michael Jordan the professor. Talking to an average person on the street and you’re going to show the greatest basketball player of all time. Retrieving the right content for the right individual under the right context is what we call personalization.

So how have we progressed to hyper-personalization? Here are the various stages of personalization that have progressed through time. Note that there is a lot of overlap in time between these various stages.

1. One-Size-Fits-All Stage: Content was generally the same for everyone. Whether it was TV programs, newspapers, or magazines, content was broadcast or printed en masse without much individual tailoring. Think content that catered to everyone like The Ed Sullivan Show or Time magazine.

2. Segmentation Stage: As more businesses began to obtain audience measurement, there was an increased understanding of the need for personalization. This began with basic segmentation - dividing users into large groups based on basic data such as age, location, gender, and perhaps browsing behavior. Content was then tailored to these broad segments. Think TV shows using audience demographics and viewing rates to determine time slots or different magazine layouts for different geographies.

3. Behavioral Personalization Stage: As data collection and analysis techniques improved, businesses began to tailor content based on individual behaviors. Recommendations on e-commerce sites or movie streaming platforms are examples of this stage, where content is tailored based on browsing and purchasing behavior, clicks, or listens. Think Netflix showing you shows you might like or Amazon recommending products for you.

4. Predictive Personalization Stage: With advancements in AI and machine learning, companies began to predict user behavior and preferences, enabling them to proactively personalize content. This stage includes dynamically personalized emails, websites, and advertisements that change in real-time based on predicted user behaviors or interests. Think Spotify creating playlists for you.

5. Hyper-Personalization Stage: This era is characterized by deep, comprehensive user profiles and a combination of real-time and predictive personalization. It leverages AI, machine learning, big data, and real-time analytics to deliver extremely personalized content and experiences. Businesses can tailor content to a degree where it seems to be created specifically for each individual user. This is just starting to happen but think of Carvana’s thank you campaign above.

Reading through the stages, the hyper-personalization stage seems like an obvious step forward. The question becomes why now? Why hasn’t it happened sooner? What is available now that wasn’t before? Let’s look at the bottleneck that was holding hyper-personalization back.

The Hyper-Personalization Bottleneck

There's three parts that are needed to effectively perform personalization.

1. Data about the individual to allow for personalization

2. Algorithms to determine what content to show a given individual

3. Content curated for an individual

Data is the first necessary component, as without it, it would be nearly impossible to personalize correctly. Entire industries are built on collecting personal data of consumers and monetizing the data. It even has a name - surveillance capitalism. Remember, if a product you are using is free, you are the product. Once they have this data, companies create segments of common groups in the data to enable the next two components.

You are likely no stranger to the various personalization algorithms you interact with on a daily basis. Recommender systems are one flavor of personalization algorithms that every major platform has implemented. These systems determine what posts show up in your Facebook and LinkedIn feeds, what videos you see on Youtube and TikTok, and select which shows pop up as recommendations on Netflix and Hulu. These algorithms work by taking the data they know about you as an individual and comparing you to other individuals on the platform and then finding content that people like you have enjoyed previously. The solution of how to find the right content for the right person has been around for decades and is well studied.

Content is what an individual ends up consuming, whether that’s a post, an image, video, audio, essay, advertisement, or Substack article. Content is the reason people use various platforms, and entire careers and industries are built around the creation of content. You might know them as writers, actors, directors, producers, musicians, advertisers, influencers, or regular people. All of these individuals create content and creating quality content is time-consuming. Due to the time component, when people talk about personalizing their content, they might make several different flavors for different groups of individuals or market segments. For instance, you might see an article about how to buy a mortgage differ if you are fresh out of college vs just starting a family vs middle aged vs in retirement. Those would be different flavors of the same article with different language and considerations.

Content creation was the bottleneck to hyper-personalization. Recently, due to generative AI, we can now create content on a massive scale. What was once a restriction to hyper-personalization has now been removed. We know how to collect massive amounts of data, how to find the right content for the right individual, and now how to create content tailored to an individual.

So, what’s the new bottleneck? Trust. How do you trust that the content that’s being automatically created is in alignment with what you want? How do you mitigate the risks of automating a creative process? How do you measure a system quantitatively that was historically measured qualitatively?

The How of Hyper-Personalization

Personalization is valuable because it has been shown to be a differentiator for top performing companies, creating as much as a 15-25% lift in revenue. Keep in mind that those numbers are based on previous methods and do not account for a world of hyper-personalization. Similar increases should be expected for early adopters who can implement and deploy hyper-personalization systems. How can we setup a system to start hyper-personalizing our content? We need the following:

1. Core piece of content

2. Hyper-targeting system and data

3. Content creation system

4. Trust layer for risk management

First, you need to start with a core piece of content. A message you want to send to the world. Great care needs to be taken here to develop this core idea as it will be translated through the personalization pipeline. Next, we need a hyper-targeting system that can determine information about a single individual. The more detailed the better as this data will be used to shape how the core content is personalized. The core content and personalization data are fed into the content creation system. At the moment, generative AI is the main engine for the content creation system. This engine is used to create millions of pieces of unique content for every individual. Finally, you need a trust layer to ensure that content is of quality, aligned with what you want, and not destructive. Perhaps you want something like BrandGuard if you are creating brand related content.

Figure 1. Hyper-personalized content creation system

Example

Working through the example in Figure 1, let’s say we run a dessert food company and we want to create an ad campaign to entice people to buy our dessert. Our plan is to hyper-personalize the image used for each individual based on what they are most likely to respond to. So, we take our core piece of content, vanilla ice cream, and combine that with all of the data our first party data system has collected about our customers. We then use a generative AI system to modify our vanilla ice cream image into an image that each individual customer is more likely to engage with. In this case, it creates a few types of desserts - more ice cream, chocolate, and flan, along with some things that aren’t dessert - pizza and hamburgers. That’s where our trust and safety layer comes in. This protective layer identifies that pizza and hamburgers aren’t in the main theme of dessert and flag those images. By regenerating those bad images, the trust and safety layer converts them to the on-target images of cupcakes and cookies. The combination of these four components allows us to automatically generate a highly targeted campaign at scale. Without all four pieces we can’t create successful hyper-personalized campaigns.

Carvana

To make the example more real, let’s step through what Carvana did. At the end of the day they walked away with a seemingly powerful hyper-personalization system.

When all was said and done, we were left with a cloud-based system capable of rendering up to 300,000 videos per hour, that’s already produced around 45,000 hours of personalized content (that’s over five years’ worth of videos watched back-to-back!).

However, If we dig in, they were able to hyper-personalize content but it was far from a fully automated system.

Creating the tech behind Joyride and its AI-generated videos was no simple task. We leveraged the expertise of hundreds of skilled professionals across dozens of disciplines.

Let’s walk through how what Carvana built fits into this framework. Their core piece of content starts with creating a video reminding the buyer about their car purchase while driving down memory lane. Next, they used information to hyper-personalize: the driver’s first name, the make and model of the car, when the car was bought, and where it was purchased. Then they triggered a process for generating content. They fed in information about what was happening in the world at the time of purchase, points of interest related to the location of the individual. The trust and safety layer was the most complex and likely not fully automated or robust. If you watch a few of the videos you can see how Carvana enforced trust in their automated system.

• Each video follows the same basic script with a few insertion points.

• The basic theme and feel of the videos are those of flowy memories to reduce the impact of any off-generated frame of the video.

• Carvana has an existing 3d model of each car. So, while each car is in the art style of the video, little error can occur with an existing 3d model.

• The name is logged in a database and is easy to show in the video. However, pronouncing the name is risky. So the narrator says “I hope I pronounced that right” and then dives into different words that start with the same letter as the name of the buyer to distract.

• The videos show two time-related components - what was celebrated that day (such as national cupcake day) and what was happening in pop culture at the time. There are only so many days in a year and major events in pop culture that each of these was likely manually reviewed before being put into the system.

• The final part of the video talks about the location of the purchase and how memories were made. While there are only so many different types of terrain and major points of interest, these are easy to verify with a combination of manual review and automated rule-based processing.

While I’m sure some models may have been used for verification, Carvana was crafty in the creation of their campaign to control for points where generative content might fail. A great start to a hyper-personalization system that will only improve with time.

Living in a Hyper-Personalized World

We’ve entered a new era that is going to change the way we interact online and in the world. By understanding each individual deeply, we can create products, services, and experiences that truly resonate with people and meet their unique needs. This is an exciting time to explore new possibilities. While the societal impacts are too early to tell, there’s a lot of cool things that can happen in a hyper-personalized world.

• Customized learning for any topic that words and stages concepts in a way that is quickest for you to understand.

• Advertisements and landing pages that are customized for a look, feel, and flow that is most comfortable to you.

• Email and news summaries that bubble out the most relevant information to you based on the current context of your life.

• Removal of excess and unnecessary information that you tend to skip over all around the web

• Preferences that span across systems. If you’re trying to learn Spanish, not only will your language app be tailored in a way to teach you the fastest, but parts of web pages can suddenly turn from English to Spanish based on your comprehension. Site layouts rearrange themselves for you to find relevant information faster while remaining unique enough to be distinct.

The benefits of hyper-personalization are immense. It holds the potential to elevate our quality of life, efficiency, and overall satisfaction. However, a lot of it is dependent on how we value privacy vs access to new technology. The key lies in using hyper-personalization responsibly, protecting individual privacy, and ensuring a balance between personalization and shared experiences for the betterment of society.