NativeAI

Content Recommendations on Digital News Websites: A Primer

By Karthik / April 5, 2019

Content Recommendations in digital media publishing started off with great intentions, but has devolved into a messy pit that publishers rush to disown.

A title that reads like clickbait, a cheesy picture and a rather poor story at the end of a link - this is what most news websites carry at the bottom and/or at the sides of their news stories, usually under the “Around the Web” or “Stories you may like” heading. These recommended stories are usually powered by an advertising platform where the race is for CTR (Click through rate) and CPM (cost per impression)

Are Content Recommendation widgets useful for digital publishers?

Several digital media publishers have traded off the value of audience attention for the somewhat meager advertising revenue from spammy recommendation widgets, while others have opted for semi-functional widgets that rely solely on freshness, unreliable tags or popularity.

This is a completely understandable dilemma, given that recommendations are a nuanced subject that need expert treatment - but they can be a terrific game-changer, when used right.

Content recommendation widgets, and the recommendation engines that power them derive their utility from the impact they create on reader behavior. If we can agree on a few simple correlations, it is easy to see why recommended content is so attractive to publishers.

Readers visit a digital news website to consume news stories, highly relevant news suggestions can induce readers to read more, out of their own accord. More reads, when combined with good stories improves reader satisfaction which can only mean good business, as long as the publisher has a proven way to monetize the traffic. So there is no question that content recommendations are useful - to both readers and publishers.

Related: Content Recommendations powered by Artificial Intelligence from NativeAI

Advantages of adding a Content Recommendation widget to your website

In 2015, the New York Times published a story on how its in-house engineering team built the next content recommender engine that powered story suggestions based on Collaborative Topic Modeling. The techniques used for recommendations have evolved since and the NYT’s recommender system has gone from strength to strength since, showing terrific growth, retention and even paywall subscriptions.        

Today, the New York Times operates a sophisticated news recommendation model, even on its email newsletters where the content is fetched dynamically, based on your reading history, preferences & interests. In the screen grab below, you can see that the stories on an email newsletter I received are populated dynamically as I received the email on April 6th, but the lead story is one that was published on April 8th.

NYT newsletter recommendations

 

Email newsletters, push notifications and other channels for personalized content recommendations can take into account a reader's reading history, remove a story that has already been read and include ones that are likely to interest her, even if it is was published after the newsletter was sent, but before it was opened.

A content recommendation engine serves a powerful purpose in shaping the traffic flow on a website, one that is increasingly critical to monetization strategies such as subscriptions and paywalls. Digital news publishers spend significant sums on acquiring new readers, and their behavior once they have read their first news story is impacted by recommendation widgets in 2 important ways:

  1. What does the reader read next?
  2. What impression does the reader form about the stories you publish?

To most seasoned news consumers, these panels become blind spots to ignore, much like banner ads. Although the advertising based recommended stories platform might claim that the stories are selected by a powerful matching algorithm, several editors and journalists view the widget with contempt due to the low quality of suggested content.

What makes Content Recommendations good?

Relevance. Analysts & Audience Development teams can come up with complex metrics and models to measure the impact of different parts of a news story that help drive recirculation, but the most important measure of success in recommendations is relevance.

How does one measure relevance, and how can publishers make relevant recommendations? The key is to build context using environment variables to segment readers and sessions.

Context for content recommendations is typically built using the following aspects and creating user segments:

  • Session attributes
    • Engagement quality, time spent on current story
    • Topics & entities covered in current story
    • Level of engagement with previous stories
    • Topics or interests covered in stories that the user read previously
    • Device used for session
    • Location & referrer for the session
    • User attributes
    • Demographics (estimate)
    • Paying subscriber or not

The above list is certainly not exhaustive and most advanced recommendation systems incorporate a multi-layered approach to personalization that uses neural networks to build comprehensive knowledge graphs of readers’ interests.

A well designed content recommendation engine understands each reader’s interests, reading history & profile features and is able to map each reader into segments. These reader segments are then associated with a set of stories that are likely to be most relevant for them to deliver appropriate suggestions.

Here’s a list of other factors that could determine Content Recommendation quality:

  • Freshness of the news story
  • Type of article - evergreen vs. temporal
  • Location of the widget (sidebar vs. end of article widget vs. newsletter)
  • Choice of featured image
  • Quality of headline text

To summarize, using a Content Recommendation engine on your digital news website can provide a terrific boost to your reader satisfaction levels, revenues and reads per session - on the condition that you pick the right content recommendation platform, deploy it across multiple channels, and configure it for sensible frequency and relevance. In some of our following posts, we will dive deeper into how you can accomplish these, but we hope this gave an overview of the benefits of using recommendations.

Written by Karthik / April 5, 2019