Washington Post seeks personalization's 'sweet spot'
A trio of ongoing projects from The Washington Post Co.’s WaPo Labs provides some valuable insights about what the future of personalized content may look like.
The three projects – Trove, Social Reader and Personal Post – offer different approaches to helping audiences discover news and other content that’s tuned to their interests. All are based on a shared architecture that combines machine-based personalization, user customization and editor curation.
“We’ve been constantly experimenting with different ways to get users into a personalized experience that feels natural to the situation,” Vijay Ravindran, SVP and chief digital officer at The Washington Post Co., said in a phone interview. “The lessons we’re learning are very much about the power of making it easy to personalize. You have to create an intuitive experience that doesn’t feel like work. We’re trying to find that sweet spot.”
Each of the three projects delivers a different user experience through a different web property. The first, Trove, a Web- and app-based service that launched in April 2011 as a collection of aggregated, public RSS feeds. Trove was WaPo Labs’ first public-facing test of personalized news aggregation using semantic content processing, and it validated WaPo’s premise that personalizing products are an important evolution of the news discovery process.
Trove provided the foundation for the September 2011 release of Social Reader, a Facebook app built with the Open Graph API that added a social layer to the personalization experience. “Where Trove.com was about users explicitly declaring topics that were of interest to them, Social Reader added algorithms for determining topic interests of users using social signals,” said Ravindran.
Social Reader also signaled a transition from Trove’s aggregation model to one featuring hosted content from the Washington Post and its syndication partners. “Because of the nature of the app, we realized the optimal user experience would only come from using hosted content,” Ravindran said. “So the platform had to evolve because now it actually had to host the content, images and video.”
As WaPo engineers layered activities from a user’s social graph onto the discovery platform, they realized they had to tweak the algorithms to give more weight to the primary user’s activities. “When we launched Social Reader, we treated the reading activity of your friends as almost equal to your own interests and reading activity,” Ravindran explained.
The problem was that Social Reader users had an average of 42 other friends also using the app – which tended to skew the personalization away from the user’s own activities. A key learning here was that machine-based personalization is neither foolproof nor sufficient by itself to provide an optimal experience.
“We realized that the platform can’t be 100 percent implicit,” Ravindran said. “Everyone will have use cases where algorithms fail you. You have to offer enough of a control panel so that if someone wants to be engaged they can tweak their own experience.”
Developers used this insight in the February rollout of Personal Post, which lets registered users on WashingtonPost.com created personalized news streams from existing site content. Rather than social signals, Personal Post uses the click history of registered users to deliver personalized content. Readers can choose to see less of a topic or remove the topic entirely from their feed.
The design meets Ravindran’s objective of a “tuning” option that doesn’t overburden users who want to customize their experience but is also easily ignored by users who want the stories selected for them.
“We’re OK with 95% of users never utilizing tuning if they don’t want to,” he explained. “But it’s there if someone really wants it. You need a strong implicit experience for the lightly engaged, but you need to give an escape outlet when they want to engage more heavily.”
Ravindran said Social Reader has 32 million “lifetime users” – meaning those who have accessed the app via Facebook Connect. The app also has more than 300,000 fans. Without offering specifics, Ravindran said these fans are consuming “quite a higher number” of articles per month than Social Reader users who have not clicked on the Fan button.
With Personal Post, deeper engagement is also the goal. Currently, users who encounter Personal Post are consuming about two more articles per visit. Ravindran said they’re also seeing an uptick in visit frequency among Personal Post users. Those are “meaningful” changes, he said.
WaPo Labs has a small team of editors who manage the three personalization platforms. “The algorithms are written by computer scientists, but they have to be tended to by people who know journalism,” said Ravindran.
Editors put their judgment to use in ways that are visible – e.g., editor’s picks – and ways that are not, such as ranking the sources of syndicated or aggregated content. Ranking sources will push what the editors see as more credible content to the top of the queue.
A New York Times article on the presidential debate, for example, would likely get preferred placement over an independent blogger’s coverage of the same event. But editors can get more granular so that lesser-known brands that have specific domain expertise rank higher than general news sources. For example, an SBNation blog on the Baltimore Ravens football team might rank higher than the New York Times’ coverage of the Ravens, even though the Times would be ranked higher for general news.
On Trove, editors’ picks appear in a news stream that’s separate from the aggregated links. On Social Reader, editors’ picks blend with the user’s personalized recommendations. “If someone is heavily engaged and has done a lot of reading or sharing activity against their topics, they might not see any editor’s picks,” Ravindran said. “But someone who’s lightly engaged will see what we call trending stories, which are editorially driven.
“When thinking about that spectrum of users, editorial judgment becomes very important for people who are lightly engaged,” he added. “Editors are helping guide content that is potentially interesting to people, and over time, personalization takes over.”
In theory, personalization aids an ad-driven revenue model by improving engagement metrics such as articles consumed and frequency of visits. More visits and more pages equal more ad revenue.
“If we can turn a user who visits three times a month into one who visits 10 times a month, that’s a big shift,” said Ravindran. “Such a shift in user behavior is possible if you build the right experience for them.” Early returns are promising, he said.
WaPo is exploring other ways to monetize its personalization efforts. For example, Trove’s semantic technology formed the basis of @MentionMachine Social Index, which tracks Twitter mentions of companies as well as topics of interest. During the Republican primaries, WaPo used the tool to track candidates’ media and Twitter mentions, selling sponsorships into a footer it ran on relevant news pages that took users to a leaderboard.
“That’s a unique use of the Trove platform that we would not have predicted,” Ravindran.
Another unanticipated development was a Trove-based ad unit, launched in April, that lets advertisers pull real-time social content into their ad creative.
“We’ve been particularly interested in using topical interests to help guide users to useful promoted content,” said Ravindran. “There’s a Holy Grail around ad units in creating a win-win situation for the advertiser and the user by delivering content that’s useful.
“If you can get to a place where you’re bookmarking a link that you were taken to from an ad because it was so useful to you, or you’re socially sharing it, that’s what we should be striving for.”