• Brett Matson

Why Content Recommenders Fail... and What to Do About It

In all my reading on digital publishing, this quote from Greg Starek at Ezoic fascinated me the most:

...getting users beyond the first and second page views offers exponential revenue potential. This kind of user engagement offers a greater return than just about any optimization in ad demand could ever provide.”

What does this mean?

His related chart of PVCount (number of page views per reader) vs. EPMV (revenue) reveals what's happening...

Revenue vs page views per session

Image Credit: Ezoic

This shows that a reader who clicks on three articles instead of one doesn’t just triple revenue, it can be a lot more. Interesting.

But why?

The report doesn’t say, so I tracked down Greg, who'd since moved on from Ezoic. He told me:

“I can only speculate, but this looked like a function of advertiser demand. Most likely causes for demand dropping are a). advertiser demand for particular users decreasing throughout a session, and/or b). url preference from advertisers. My best guess is this is a combo of both.”

Either way, there’s a lot to be gained from encouraging audiences to consume additional articles. And this is the reason for the growing content recommendation engine industry.

The Problem with Content Recommenders

In recent years, recommenders have evolved to almost exclusively employ machine-learning algorithms such as collaborative filtering to identify content that is most likely to appeal to a particular reader based on what’s known about them. It’s the Netflix method of personalising by recommending TV shows viewed by people who viewed shows similar to the ones you’ve viewed.

It works for Netflix, so does it work for digital publishers?

Not as much - let's look at the reasons why.

Lack of User Data

Netflix has a vast amount of user behaviour data to drive its recommendations, but how useful can a recommender be for a reader visiting a digital publisher site for the first time?

This problem is glaring when it happens on product recommendations. For example, JB Hi-FI is currently selling the newly-released 65” 8k Neo QLED Smart TV for $4,795 and it’s product recommender suggests that this new-model TV is frequently bought together with Ella Fitzgerald on vinyl. I wonder how many customers actually bought that combination?

Ad for Samsung New QLED TV for $4,795
Recommendation for Ella Fitzgerald on vinyl

A Shot in the Dark

Another problem with content recommenders is that with limited screen real estate, they can’t recommend more than a handful of articles. Combine this with a lack of user behaviour data and providing recommendations becomes a shot in the dark for most publishers.

Coming Late to the Party

Finally, an overlooked problem with content recommenders is that they attempt to do their thing after the reader has disengaged from the article. It doesn't matter how accurate the recommendations are if the browser tab is closed before the reader scrolls onto them.

Is there a Solution?

Traditional machine-learning based recommenders can be an effective tool when done right, but it’s likely they’ll be leaving publisher revenue on the table.

The Engaging Things platform is designed to solve the limitations of traditional recommenders by:

  • Providing recommendations while the reader is still engaged in the article.

  • Showing a much broader range of related content.

  • Organising large numbers of recommendations into easy-to-browse categories.

  • Working effectively when no user behaviour data exists.

Feel free to get in touch to find out how this could enhance your content revenue.