Building better metrics for news

Building better metrics for news

"Pageviews are dead"

Remind you of anything?

Pageviews Above Replacement (un-juking the stats)

• What if we could control for promotion when judging performance?
• From July - August, I collected data on the promotion and performance of over 21,000 articles published on nytimes.com

Data sources

• Promotional Data:
• ~ 200 NYT-related Twitter accounts
• ~ 20 NYT-related Facebook accounts
• ~ 20 section fronts
• One homepage
• One paper Metadata:
• Article type: (video, slideshow, interactive, article, blogpost)
• Section: (US, World, Art, etc...) Performance Data:
• Pageviews and Social Media Activity for each article

Predicting pageviews

• Sum all the pageviews for 7 days on the site
• Use promotional features and article metadata to predict this number
• Random Forests (the mode of a bunch of decision trees)

Variable importance

• Time on all section fronts
• Number of unique section fronts
• Was the article in the paper?
• Number of NYT-Twitter followers reached
• Time on homepage
• Number of NYT-tweets
• Is the article from Reuters?
• Is the article from the AP?
• Max rank on homepage
• Word count

So what?

• Placing promotional data alongside pageviews gives us a better understanding of what the metric actually means.
• (NYT) Pageviews are actually fairly predictable (90% of the variance explained in my model)
• Incorporating this approach in your Newsroom should be fairly painless with particle. However, you should first ask yourself what you're optimizing for.
• Predictive analytics can help increase your editorial responsiveness to the reader's preferences, the news cycle - http://fast.qcri.org/.

Thanks!