This post is the third in a series on the eContext API and how our clients get value out of semantic classification. To learn more about our technology and its uses, feel free to get in touch.
In the previous installment of “The Adventures of the eContext API User” (working title), I touched on different ways businesses can better understand their audiences’ interests—self-disclosed through billions of digital interactions—and discussed a few ways to profit from that insight.
I have to admit, that post was a little bit guilty of a flaw that’s typical in conversations about data: the straightforward argument that “more data = more relevance = happier users.” While this is generally true, especially when catering to entire audiences, things get a little more complicated when you talk about individuals. Today, we’re going to talk about single user profiles—not just how to divine a consumer’s interests, but how to leverage that information with care.
For the sake of clarity, let’s discuss this issue using a typical eContext use case:
- The client is media group with several digital magazines, mainly offering articles and video content on a wide variety of different topics
- The client’s primary source of revenue comes from selling ad impressions
- Client’s goal is to facilitate better individual personalization—tailoring recommended content to a user’s demonstrated interests—in order to cross-promote sibling properties and keep the user on client-owned websites for longer browsing sessions
Step One: Classify Existing Content
If you’ve seen any of our previous posts, you’ll probably be familiar with this part: our client will pass their content through the eContext API and receive precise, structured data on the topics covered by that content.
Here are the different API calls the client would use for different kinds of content:
- The URL of every client-owned webpage will be submitted via Classify/url, giving the client a list of the most prevalent topics on each page.
- The client’s properties feature comment sections at the end every article; this user-generated content is submitted via Classify/social.
- For video content, the client has a few different options:
- Relying on Classify/url to classify the elements surrounding the video: the title, description, etc.
- If the client has a reliable transcript of a video, that transcript can be passed through either Classify/social or Classify/text.
- If the client can’t get a reliable transcript AND classifying the video-adjacent elements is insufficient, eContext can still help. Because we operate our own transcription models, we can process and classify video content in-house. Note that this is not an API call, but an agreed-upon project to be carried out by eContext.
The client will then metatag every page, article, video, and message board post with eContext’s topic classifications. Remember: our topics are nested within a giant subject matter hierarchy. That means an article comparing the best brands of hair spray won’t just be labeled with Hair Spray, but also Beauty, Hair Care, Hair Care Products, and Styling Products. This hierarchical structure is going to come in handy, as I’ll discuss later in this post.
Step Two: Classify New Content, Search Queries
Obviously, the client can’t just call it a day after labeling all of its existing content. New material is being published daily, so the client sets up real-time access to Classify/url and Classify/social in order to stay up-to-date.
In addition, the client will use Classify/keywords to label the topic of each search conducted across its properties. Remember, the client is using these classifications to both learn about users and deliver content that matches their interests, so keeping track of what a user is searching for will be a key source of information (especially if the user is searching for a topic that client doesn’t have content about!)
Step Three: Building Profiles
Most sophisticated digital properties have a method of keeping track of users, whether through cookies, an IP address, a site-specific user account, or social media authentication. The important thing is that the client can keep track of all the content consumed by a user over time, any searches he or she makes, and the messages he or she posts to the website’s comment sections.
Because all of those artifacts have now been classified to topics thanks to eContext, the user’s activity yields an active, real-time profile that communicates the user’s interests. Here’s a basic visual representation of what such a profile might look like:
Now, there are a lot of different ways to rank what a user is “most” interested in. How to weigh the topics of a user’s searches against the topic he or she ends up reading about? Should rankings decay over time to reflect a user’s developing interests? If a user spends an above-average time on a particular article, should that article’s topics be prioritized? These are questions best answered on a case-by-case basis per the goals of an individual client. However, feel free to contact us if you ever want to go over a few recommended practices.
Before Step Four, Let’s Talk about Relevance
I’m a pretty big nerd, no doubt about it. If I’m on Buzzfeed and I see a link on the homepage to “The Best Easter Eggs in ‘Stranger Things’ ”, there’s a pretty good chance I’ll click.
On the other hand, if I see “The Best Easter Eggs in ‘An Age of Kings’ ”, I’ll probably be a little creeped out. That’s because An Age of Kings isn’t some hot new Netflix show; it’s a BBC Shakespeare adaptation from 1960 that’s only available on DVD, and the only reason I would see a link like that is if the website somehow knew that I’d recently purchased it.
All of the different little personalizations in our digital lives—suggested purchases, your home marked on Google Maps, that video Facebook makes for your friendiversaries—are a reflection of who you are in the real world. As that reflection achieves more and more granularity, it approaches a kind of online uncanny valley; it’s so accurate that it starts to get disconcerting.
Here’s my point. eContext gives clients the opportunity to understand users with unparalleled precision, but the real challenge is finding the right balance between broad and granular personalization.
High-Fidelity Personalization
Pro: Guarantees relevance
I love it when Amazon knows exactly what I’m looking for!
Con: Potentially unsettling
Why am I seeing ads for maternity wear?! We haven’t even told our friends!
Low-Fidelity Personalization
Pro: Yields suggestions that might not be 100 percent relevant, but are close in a way that can broaden existing interests
Spotify’s Discover Weekly playlists are awesome… I never knew I liked EDM!
Con: Increased risk of unhelpful recommendations
I bought some measuring cups last week, so now the internet thinks I’m a bakery.
With this delicate give and take and mind, we’ll now consider how our client can leverage the user profiles they’ve obtained thanks to eContext.
Step Four: Delivery, Feedback, and Calibration
Across its digital magazines, the client can now take stock of a visitor’s interests and offer up content to match. More than that though, thanks to the hierarchical nature of eContext topics, the client is free to calibrate exactly how precisely to offer up those recommendations.
Just because the user has shown an interest in Hair Spray doesn’t mean that the client must restrict recommendations to that topic. That could get old, redundant, or even a little too close for comfort. If Hair Spray suggestions aren’t getting traction, the client may set its recommendation algorithms to consider the broader categories like Hair Care or even Beauty. By using eContext’s broad-to-granular topic structure and consistently calibrating towards the right level of specificity, the client can continually flex its personalization strategy and optimize visitors’ time-on-site.