Predicting Content Attention and Behavior

Audiences, at times, seem inscrutable. We want to know how audiences will respond to content, but audiences don’t behave consistently.  Sometimes they skim content; sometimes they read it closely.  Even if we wish it were otherwise, we can’t escape the messy reality that audiences in many respects don’t behave in a single consistent way.  But do they behave predictably?  Can we predict want kind of content will engage online audiences, if we could account for known variables? To date, progress untangling this problem has been limited.  But we have reason to be optimistic it won’t always be this way.  A more data-centric approach to content strategy could help us understand what variables influence audience behavior.  

The biggest weakness in content strategy today is that it lacks predictive explanatory power.   Whenever someone advances a proposition about what audiences want or will do, it is easy to find counter examples of when that doesn’t hold.  Nearly all categorical assertions about what people want from content don’t survive even minimal scrutiny.  Do audiences want more or less content?  Do they want simple or detailed explanations?  Do they want to engage with the content, or get the content in the most digestible form possible? Such binary questions seem reasonable to ask, and call for reasonable  answers in return.  But such questions too often prompt simplistic answers that promise a false sense of certainty.  Content behavior is complex — just like human behavior in general.  Yet that doesn’t mean it is not possible to learn some deeper truths — truths that may not be complete and exhaustive, but are nonetheless accurate and robust.  What we need is better data that can explain complexity.  

To provide predictive explanatory power, content strategy guidelines should be based on empirical data that can be reproduced by others.  Guidelines should be based on data that covers a breadth of situations, and has a depth of description.  That’s why I was so excited to read the new study presented last week at 2018 World Wide Web Conference by Nir Grinberg of Northeastern University, entitled “Identifying Modes of User Engagement with Online News and Their Relationship to Information Gain in Text.”  The research provides a rare large scale empirical analysis of content, which reveals many hidden dimensions that will be useful to apply and build on.  I encourage you to read the study, though I caution that the study at times can be dense, filled with academic and statistical terminology.  I will summarize some of its highlights, and how they can be useful to content strategy practitioners.  

Grinbert’s study looked at “a large, client-side log dataset of over 7.7 million page views (including both mobile and non-mobile devices) of 66,821 news articles from seven popular news publishers.”   By looking at content on such a large scale (nearly 8 million page views), we can transcend the quirks of the content we deal with in our own projects.  We want  to understand if the features of our content are typical of content generally, or are characteristics that apply to only some kinds of content.

The study focused content from news websites that specialize in different topics.  It does not represent the full spectrum of content that professionals in content strategy address, but it does cover a range of genres than are commonly discussed.  The study covered seven distinct genres:

• Financial news
• Technology
• How To
• Science
• Women
• Sports
• Magazine features

Grinbert was motivated by a desire to improve the value of content.  “Post-hoc examination of the extent to which readers engaged with articles can enable editors to better understand their audience interests, and inform both the coverage and writing style of future articles.”

Why do analytics matter? Content that audiences use is content that audiences value.  The question is how to measure the audience use of content, after they click on a link.  Page views are not a meaningful metric, since many views “bounce”.  Other metrics draw controversy.  Is a long time on a page desirable or not desirable?  With simple metrics, the metric can become hostage to one’s own ideological worldview about what’s best for users, instead being a resource to learn what users are really trying to accomplish.

First, how can we measure attention?  The study considered six metrics available in analytics relating to attention:

1. Depth — how far scrolled in an article, a proxy for how much of the content was viewed or read
2. Dwell time — total user time on a page (good for non-reading engagement such as watching a video)
3. Engagement — how much interaction happens on a page (for example, cursor movements, highlighting)
4. Relative depth — how much of an article was visible on a user’s screen
5. Speed — speed of scrolling,  a proxy of how quickly the readers “read” the content
6. Normalized engagement — engagement relative to article length

The metrics that are “relative” and “normalized” attempt to control for differences between the absolute values of shorter and longer content.  

Next, what might these metrics say about audience behavior?  Through a cluster analysis, the study found these indicators interact to form five content engagement patterns:

• Shallow  (not getting far in an article)
• Idle (short period of activity followed by period of inactivity, followed by more activity)
• Scan (skimming an article quickly)
• Read (reading the article for comprehension)
• Long read (engaging with supplementary materials such as comments)

So how do specific behaviors relate to engagement patterns?  The study showed that the indictors were associated with specific engagement patterns.

Depth  (ranked from low to high depth of scrolling)

1. Shallow
2. Idle
3. Scan
4. Read
5. Long read

Dwell time (ranked from short to long dwell time)

1. Scan
2. Read 
3. Long read 
4. Ide
5. Shallow

Engagement (ranked to low to high engagement)

1. Shallow
2. Scan
3. Idle
4. Read
5. Long read

Relative depth (ranked for short to long relative depth)

1. Shallow
2. Idle
3. Scan
4. Read
5. Long read

Speed (ranked from fast to slow)

1. Scan
2. Read
3. Long read
4. Idle
5. Shallow

Normalized engagement (ranked from low to high)

1. Shallow
2. Idle
3. Scan
4. Read
5. Long read

So what does this mean for different kinds of content? “We found substantially more scanning in Sports, more idling in “How To”, and more extensive reading for long-form magazine content.”  That may not sound like a profound conclusion, but it feels valid, and it’s backed by real world data. This gives us markers to plan with.  We have patterns to compare. Is your content more like sports, a how-to, or a feature?   

For sports, readers scan, often just check scores or other highlights, rather than read the full text.  They are looking for some specific information, rather than a complete explanation.  Sports is a genre is closely associated with scanning.  When sports is presented in long form, such as done on the now defunct Grantland website, it only appeals to a niche.  ESPN found Grantland unprofitable.  Grantland violated the expectations of the genre.  

Magazines were most likely to be read shallowly, where only the first few sentences are read, as well as the most likely to be read thoroughly, where even comments are read.  This shows that the reader makes investment decision about whether the content looks sufficiently interesting to read in depth.  They may leave a tab open, hoping to get back to the article, but never doing so.  But sometimes, a preview summary such as an abstract can provide sufficient detail for most people, and only some will want to read the entire text.   

The study found  a “relatively high percent of Idle engagements in How To articles. The few articles we examined from this site gave instructions for fixing, making, or doing something in the physical world. It is therefore plausible that people disengage from their digital devices to follow instructions in the physical world.”  

How the Study Advances our Practice

The study considers how reading characteristics converge into common reading patterns, and how different genres are related to distinct reading patterns.  

The study brings more a sophisticated use of metrics to infer content attention.  It shows how features of content influence attention and behavior.  For example “total dwell time on a page is associated with longer articles, having more images and videos.”   Not all content is text.  How to measure use of video or images, or exploring data, are important considerations.

We have concrete parameters to define engagement patterns.  We may casually talk about skimming, but what does that mean exactly?  Once we define it and have a way to measure it, we can test whether content is skim-able, and compare it to less skim-able content.  

Solid detailed data helps us separate what is happening from why it may be happening.  Slow reading speed is necessarily not an indication that they material is difficult to read.  Fast reading speed doesn’t necessarily indicate the topic is boring. Readers may be involved with other activities.  They may have different knowledge already that allows them to skim.  Instead of debating what is happening, we can focus on the more interesting topic of why it might be happening, and how to address it.   And with benchmark data, teams can test alternative content designs and see how the performance changes.  

How Content Strategy can build on the study

The study shows that more robust analytics can allow us to directly compare utilization characteristics of content from different sources, and compare the utilization characteristics of different genres and formats of content.  Standardized data allows for comparisons.

The study suggests more sophisticated ways to measure attention, and suggests that attention patterns can depend on the genre of content.  It also identified six content behaviors that could be useful for classifying content utilization.  These elements could contribute to a more rigorous approach to using analytics to assess audience content needs.

A framework using detail metrics and patterns can help use baseline what’s actually happening, and compare it with what might be desirable.

For example, what kinds of content elicit shallow engagement?  Is shallow engagement ever a good, or at least an opportunity?  Perhaps people start then abandon an article because it is the wrong time for them to view it. Maybe they’d benefit from a “save for later” feature.  Or alternatively, maybe the topic is valuable, but the content is uninviting, which grinds the engagement to a halt.  With more a sophisticated ability to describe content behavior, we can consider alternative explanations and scenarios.  

The study also opens up the issue of whether content should conform to typical behavior, or whether content should try to encourage a more efficient behavior.  If How To content involves idle periods, should the content be designed so that people can start and stop reading it easily?  Or should the content be designed so that the viewer knows everything they need to do before they begin (perhaps by watching a video that drills how to do the critical steps), so they can complete the task without interruption?   I’m sure many people already have opinions about this issue.   More precise analytics can allow those opinions to become testable hypotheses.

The big opportunity is the ability to compare data between content professionals, something that’s not possible with qualitative feedback.    Today, we have conferences where different people present case studies.  But it is hard to compare the learnings of these case studies because there are no common metrics to compare.  Cases studies can also be hard to generalize, because they tend to focus on process rather than focus on common features of content.  Two people can follow the same process, but have different outcomes, if the features of their content are different.  

Like the field of usability, content strategy has the opportunity to build a set of evidence-based best practices.  For example, how does having a summary paragraph at the start of an article influence whether the entire article is read?  Different content professionals, looking at different content topics, could each test such a question and compare their results.  That could lead to evidence-backed advice concerning how audiences will likely react to a summary.  

The first step toward realizing this vision is having standard protocols for common analytics tools like Google Analytics, so that different website data are comparable.  It’s a fascinating opportunity for someone in the content strategy community to move forward with.  I’m too deep in the field of metadata to be able to work on it myself, but I hope others will become interested in development of a common analytics framework .

— Michael Andrews