Twitter’s New Order
Inside the changes that could save its business—and reshape civil discourse.
What do you see when you open Twitter? Until a year ago, the answer was straightforward: With minor exceptions, you’d see every tweet from every person you followed, in chronological order, with the most recent at the top.
In February 2016, word leaked to BuzzFeed that Twitter was planning a move that would change everything. The company was introducing what insiders called an “algorithmic timeline.” It meant that tweets would no longer appear in the order they were posted. Instead, a complex, opaque software program would decide which tweets you’d see when you opened the app.
It sounded nefarious. Worse: It sounded like Facebook, the older, more mainstream social network that Twitter’s cool kids shun. Longtime users revolted, channeling their indignation into a bitter hashtag: #RIPTwitter. Meanwhile, a handful of techies, investors, and contrarian pundits countered the gloom with sunny predictions. An algorithm, they argued, was just what Twitter needed to reverse its fortunes and join Facebook in the ranks of social media giants.
But a funny thing happened when Twitter launched the new feature a month later: nothing. So it seemed, at any rate, to casual users of the service. Implemented in a surprisingly modest form, the algorithm so far has neither saved nor killed Twitter but wrought effects so subtle that they’ve passed almost without mention. After all the outrage, fewer than 2 percent of all users opted out of the algorithmic timeline.
Yet the changes to Twitter’s underlying structure run much deeper than outsiders realize. At a critical moment in the company’s history—and that of our body politic—the algorithm is quietly starting to reshape both Twitter’s business and the way people experience it. That includes the president of the United States, his 25 million followers, the activists opposing him, and the media that must make sense of it all.
For this article, Twitter offered a glimpse into the workings (and continued evolution) of the algorithm for the first time since it was launched a year ago. On the most immediate level, the new timeline has clearly made the service a little friendlier and livelier. It ensures that you see more tweets from the people you interact with the most and more of the most popular tweets from others you follow. It has also ensured that the most popular tweets are far more widely seen than they used to be, enabling them to go viral on an unprecedented scale.
The company says the effect has been to draw in new users and make the old ones more active. In a time of crisis for Twitter’s business, it has driven desperately needed increases in key metrics such as monthly active users, impressions, and time spent on the site. Those gains have yet to reverse the company’s overall slump, but they offer a beam of hope amid the gloom—especially since Twitter is only beginning to tap the algorithm’s potential.
But you can’t see more of some kinds of tweets without seeing less of others, and the hidden consequences of that equation could affect us all. As it gradually tightens the loops in Twitter’s social fabric, the algorithm risks further insulating its users from people whose viewpoints run counter to their own—a phenomenon, already rampant on Facebook, that has contributed to the polarization of the American electorate and the Balkanization of its media.
Facebook has taken the brunt of the blame for the fake news and sensationalism that polluted political news in the 2016 U.S. presidential election, both because it is bigger and because its more potent algorithm lends itself to those pitfalls. But Twitter played a role, too, and with the world’s most powerful person setting national policy via tweets on a daily basis, the service has never been more influential than it is today.
The question now for Twitter is whether a service that doubles as a global news ticker and water cooler can seize this moment to regain its business footing. The rest of us need to ask a different question: Should we hope that Twitter succeeds?
* * *
Despite insisting that the algorithm is working, Twitter declined to share its precise impact on key metrics with me. This reticence suggests that the effects, while positive, are not yet impressive enough to comfort the company’s restless investors. Its latest earnings report, covering the fourth quarter of 2016, showed modest gains in active users and engagement, which CEO Jack Dorsey attributed to “better relevance in both the timeline and notifications.” Yet the company’s revenue flat-lined, and its stock dropped.
Twitter’s engineers are constantly probing how well the service is engaging its users, running tests that quietly enable tweaks for a small fraction of Twitter accounts, then studying the effects on their behavior. Those tests generate much richer insights now that Twitter can toy with the ordering of users’ feeds. “Everything we are doing, we are measuring if it’s working or not,” said Deepak Rao, the product manager who oversees the company’s home timeline. “We run dozens of experiments every month.”
I spoke in depth with Rao to better understand this process and the thinking behind Twitter’s algorithm. In our conversation, Rao described a system that is still in its nascent stages yet is already far more complex and subtle than most users realize. It is a system so finely personalized that no two users will experience it in the same way yet rudimentary enough that its engineers are still struggling to ensure that it doesn’t show you the same people’s tweets every time you open the app.
The company told me that its own data show that the algorithm has boosted users’ engagement along every major yardstick it watches. Not only are people spending more time reading, favoriting, and retweeting as a result, but they’re actually tweeting more themselves—an outcome that surprised Twitter’s own product managers. “Every possible engagement and attention metric went up” when the algorithm took effect last year, Rao told me. By exactly how much, the company declined to disclose, although a spokesperson called it “one of our most impactful product launches.”
As a result of the experiments made possible by the algorithm, Twitter knows more about its users than it ever did before, such as how much they value recency or how they react to seeing multiple tweets in a row from the same person. The company has tried out new features that group tweets about a given topic or hashtag within your feed. It has even experimented with showing you occasional tweets from people you don’t follow, if Twitter’s ranking system shows that you’re likely to want to see them. Twitter can now evaluate the efficacy of such new features by comparing their effects on user behavior to the effects of the ranked timeline and “In case you missed it,” another newish feature. “Our algorithm changes on an almost daily to weekly basis,” Rao said.
All these tweaks have yet to bear much fruit, from investors’ perspectives. The good news for the company is that, when managed properly, machine-learning algorithms can be radically improved with time. Likewise, upticks in user engagement have a way of gaining momentum, as engagement begets engagement. Muted as its impact has been to the business so far, the algorithm may still power future product changes that both lure and retain new users—and, ultimately, get Twitter growing again in a meaningful way.
* * *
Twitter is sometimes accused by its loyalists of making changes too rashly. In fact, the company has treated its core product with excessive caution. The failure of former CEO Dick Costolo, pushed aside in 2015, was not that he ruined the product, as so many feared. It was that he accorded it too much reverence. The business evolved on his watch, but the user experience stagnated. The timeline when he left looked and functioned much the same as it did when Costolo arrived in 2010.
His successor, Dorsey, approved the timeline algorithm in 2015, and its success or failure will probably be attributed to him. But the minds who developed and championed the idea within the company included former CTO Adam Messinger, former head of engineering Alex Roetter, and former head of product Kevin Weil. (That all three have since left is a sign of the long-running dysfunction atop Twitter’s org chart.)
To understand why they saw the algorithm as vital to Twitter’s future, it helps to recall what preceded it. The reverse-chronological timeline stemmed from the site’s origins as a way to blast brief, real-time “status updates” via text message to friends and acquaintances. But over the years Twitter morphed into something more like a public platform for news, opinions, jokes. As the user base and its follow lists grew, the chronological feed’s limitations became clear. You’d log in and find yourself thrust into the middle of dozens of unrelated, often insider-y conversations, and the good stuff required tedious scrolling to unearth. For the ordinary internet user, it simply wasn’t worth the trouble.
This led to Twitter’s most existential and enduring problem as a business: its inability to retain a large proportion of the new users who sign up. In December 2012, Twitter announced that it had 200 million monthly active users. CEO Costolo predicted that Twitter would hit 400 million within a year. Instead, the company filed to go public in late 2013 with only 218 million monthly active users. More than three years later, that active-user number is still just 319 million, and growth has slowed to a trickle. Though the comparison isn’t perfect, the company plateaued at roughly the age at which Facebook took off. A crucial difference between the two services: Facebook’s news feed algorithm, which the company implemented early on and has been aggressively improving ever since. Whereas right now Twitter’s algorithm affects only the tweets at the very top of your feed, Facebook’s automatically orders every post according to a highly sophisticated formula that is personalized to each user’s habits, tastes, and relationships.
Part of Twitter’s problem has been its struggle to define the timeline’s precise purpose. Rao told me that this has become clearer since the algorithm launched, that the company now sees the timeline’s function as “helping users to stay informed with what’s going on in the world.” Twitter, in other words, is no longer a social network, at least by its own reckoning. It’s a real-time, personalized news service. And since there are no human editors, it falls to Twitter’s algorithm to determine which tweets will lead the news each time you open it.
Of all Twitter’s efforts to address its retention problem—and there have been many—the algorithm is its boldest. But what is the Twitter algorithm, and how does it work? The short version is that it’s a software program that evaluates tweets according to various criteria, then chooses a handful to show each user at the top of his or her timeline upon opening the app. The rest of the timeline remains reverse-chronological, at least for now. For the curious (and the confused), here is the long version—the first public peek into the algorithm’s workings.
As soon as you open it, Twitter quickly collects and assesses every recent tweet from every person you follow and assigns each one a relevance score. This score is based on a wide array of factors, ranging from the number of favorites and retweets it received to how often you’ve engaged with its author lately.
At the same time, the algorithm is assessing a variety of other variables—including how long you’ve been away from the site, how many people you follow, and your individual Twitter habits—to determine exactly how those scores will affect what you see in your feed. (All of this happens in the background.)
The algorithm’s output can take different forms in your feed, but the “ranked timeline” and “In case you missed it” are the most notable additions. The ranked timeline is what was supposed to herald the end of Twitter as we know it. Visit the site or open the app after a few hours away, and the top of your feed will look much the same as it did a year ago, with a series of tweets listed in reverse-chronological order. But examine the time stamps and you’ll notice that these tweets aren’t quite as recent as you might expect. The top one might have been posted 10 or 15 minutes ago. Scroll just a few tweets down, and you might see something published an hour ago or more. Together, these are the tweets that Twitter’s algorithm has “ranked” for you to see first.
While Twitter won’t disclose all the signals involved in the ranking—“thousands,” a spokesperson told me—the company did specify a few of them. They include:
- A tweet’s overall engagement, including retweets, clicks, favorites, and time spent reading it
- A tweet’s engagement relative to other tweets by the same author
- How recently the tweet was published
- How often you engage with the tweet’s author
- How much time you spend reading tweets by that author, even if you don’t engage.
- What kind of attachment the tweet includes (e.g. link, image, video, none), and what kind of attachments you tend to engage with.
The tweets that appear in this ranked section of your timeline constitute a tiny subset of the tweets you missed since the last time you were active on Twitter. So if you keep scrolling, soon enough you’ll reach a tweet that was published even more recently than the one that appeared at the very top of your feed. From that point on, your Twitter is back to normal, displaying every tweet from every person you follow in reverse-chronological order. And when you refresh your timeline, the ranked tweets will sink out of view.
That the ranked tweets are easy to miss, Rao told me, is by design. The company has tried at several junctures in its history to cordon off a batch of algorithmically selected tweets from the rest of the timeline. But these efforts were mostly stumbles because they didn’t feel like part of the core experience. (Remember the #Discover tab? Neither do most people.) They were “too module-y,” in Rao’s words. He said the company ultimately decided to incorporate the ranked tweets directly into the timeline so that they wouldn’t detract from the “liveliness” of the Twitter experience.
The ICYMI feature, which used to be called “While you were away,” predates the ranked timeline and remains under a separate label from the rest of your feed. Given its move away from self-contained modules, many assumed Twitter would dispense with ICYMI when the ranked timeline was launched, but the company has retained it as a complementary feature. Whereas the ranked timeline will appear at the top of your feed after just an hour or two away, ICYMI typically enters your feed only when it’s been several hours or a few days since you last opened Twitter. The tweets that appear there are less recent, and they don’t appear in chronological order at all. Rather, they’re ordered according to their ranking scores. The tweet at the top of your ICYMI box, then, will be the one that ranked highest out of all the possible tweets from everyone you follow since the last time you logged on. It’s Twitter’s equivalent of the top post in your Facebook feed.
From a design standpoint, Twitter’s use of both the ICYMI box and the ranked timeline is a clunky arrangement. But if people find it confusing, Rao said no one has told him so. The goal, in his mind, is for users not to have to think about which set of ranked tweets they see at what time, or how many, or why. Twitter’s algorithm is supposed to do that thinking for them.
* * *
Twitter, the technology and media writer John Herrman once wrote, is a truth machine. The network’s inherently public structure made it relatively easy to debunk the sorts of viral rumors and misinformation that tend to spread unchecked on Facebook. But as Twitter gets better at showing users the tweets that most resonate with them, the risk is that it’s also getting better at reinforcing their biases and abetting their construction of alternate realities—not a marketplace of ideas, but a battlefield pocked with foxholes. This past election cycle, those searching for truth on Twitter could still find it, but for others the service doubled as a lie machine—a place where falsehoods and fake news flourished among isolated ideological subcommunities that dwell in divergent realities. The same social network that had helped to call the world’s attention to Tahrir Square and Ferguson became a breeding ground for conspiracy theories such as Pizzagate. The same service that gave Trump’s critics a platform to counter his rhetoric with facts also gave his supporters the power to drown them out in a cacophony of abuse and invective.
The reason for this self-reinforcing dynamic is unsurprising: When you draw on users’ past habits to shape their future experiences, you risk enclosing them in bubbles of their own making—what Eli Pariser called a “filter bubble.” The term has often been applied to Facebook, due to its heavy reliance on a personalization algorithm that weights each post according to, among other metrics, how likely you are to hit “like.” Old-school Twitter could be a bubble of sorts, too, depending on who you followed. But the chronological timeline at least gave equal weight to every tweet, regardless of whether it was likely to please or upset you.
The ranked timeline, even in its modest present form, has changed that. You’re now far more likely to see certain types of tweets than others when you log in. The question is: What types of tweets are you seeing more of?
While social-media ranking algorithms are incredibly difficult to perfect, it actually isn’t that hard to improve on a purely chronological approach when it comes to generating engagement. Without an algorithm, users might log in and see at the top of their timeline a random thought from technology writer Farhad Manjoo that got far less engagement than his typical tweets. But even a rudimentary ranking system could ensure that users are more likely to see a viral joke from comedy writer Dan Amira, playing off a popular meme, which generated far more engagement.
The question, though, is how an algorithm alters the overall playing field—whose tweets tend to flourish, and whose wither on the vine. One person who has almost certainly benefited from Twitter’s ranked timeline is Donald Trump. Trump may not have won the presidency because of Twitter, but it’s hard to imagine his campaign strategy succeeding without it. Mocked and disdained by the mainstream media, Trump used the platform as his megaphone, bypassing editorial filters to address voters directly, in his own words—and with his own facts. With each tweet gaining tens of thousands of shares, and some far more than that, Trump’s Twitter account became a major media organ in its own right, helping to dictate the political news of the day. While no data are available on exactly what role the algorithm played, it’s a safe bet that Trump’s tweets regularly topped his followers’ ranked timelines, ensuring that the missives reached a much wider audience than they would have under the old system.
And it’s not just his fans. Before the algorithm, I used to see Trump’s tweets only when he happened to publish while I was online or shortly beforehand. They would trigger a sudden cavalcade of retweets and commentary in my feed, but it would quickly die down. Now the president’s tweets—in their original, unfiltered form—appear routinely near the top of my feed, even if he published them hours ago. I see at least one of them most weekday mornings when I log in as I prepare to board the subway to work. It’s often joined there by one or two of the cleverer or more trenchant responses from people I follow. (Twitter has also recently begun using an algorithm to order the replies to popular tweets, giving rise to a cottage industry of “first replies” that reach a substantial portion of the original tweet’s audience.)
Has this improved my Twitter experience? On the whole, I’d argue it has. Trump’s tweets and the commentary around them, for better or worse, are part of what I come for now. Twitter’s algorithm has successfully detected that, even though I almost never favorite them myself. (It may be because I occasionally quote-tweet or reply to them, or it may be because so many other people on Twitter interact with them.) By making sure I see Trump’s tweets without having to seek them out, Twitter’s timeline software is doing the job Rao asked of it. It’s telling me what’s going on in the world, or at least that portion of the world that generally concerns me.
On the other hand, the commentary it shows me about Trump’s tweets—and about politics in general—almost always comes from the left. No doubt that’s largely a function of the people I’ve chosen to follow: Most of them are liberal. Yet I’ve also taken care over the years to follow a number of pundits whose politics I disagree with. They tend to skew toward the moderate, #NeverTrump side of the spectrum, although some do support the president. Their tweets often irk me, sometimes upset me, and occasionally infuriate me. But I’ve always continued to follow them because it’s important to me that my Twitter feed not insulate me altogether from opposing viewpoints. I rarely favorite or rebroadcast their tweets, or even click their links. But I do read them, and on the whole I find them indispensable.
Interestingly, Twitter’s timeline algorithm seems not to have picked up on this. For whatever reason, conservatives’ tweets virtually never seem to crack my ranked timeline or my ICYMI box.
This has implications far beyond my own attempts to consume a balanced media diet. If you’re a right-winger who watches Fox News and reads Breitbart, you might still follow a handful of mainstream news outlets on Twitter. But if you tend not to favorite or retweet their tweets, Twitter’s software could decide that you don’t really care to see them in your feed after all. To boost your engagement, it might instead serve you ever more tweets from the same few people whose tweets you favorite and retweet the most. Within that subset, it might further emphasize the tweets that are getting most widely favorited and retweeted by others who already think like you do. Stoking a few Pizzagate embers might cause the conspiracy theory to flourish on your feed.
The good news is that Twitter is not in denial about them, as Facebook CEO Mark Zuckerberg seemed to be about his own platform’s shortcomings in the election’s immediate wake. Twitter assures me that it’s both concerned by and actively working to mitigate the algorithm’s potential to reinforce biases.
Taking more into consideration than retweets and favorites is part of this effort. A Twitter spokesperson explained to me that a Bay Area resident might never interact with the Caltrain account she follows for its service updates. But Twitter’s software can learn over time that she tends to stop scrolling long enough to read a Caltrain tweet before moving on. It could then make sure to show her a noteworthy Caltrain update in her ranked timeline during commute hours.
Rao told me Twitter also runs qualitative surveys meant to complement its data on users’ behavior, to help distinguish between what the team thinks of as “salad” (like Caltrain’s tweets) and “doughnuts” (such as an ideological exhortation that plays to your preconceptions). Still, it’s clear to anyone who uses Twitter regularly that the viral “doughnut” tweets now go more viral than before and that you tend to see more tweets from the people whose accounts you engage with most frequently. In our polarized political climate, the potential implications could be dire—not because of Twitter’s sheer scale, but because the people in power in Washington, Hollywood, and the media are among those who use it the most.
Rao said the company has noticed this homogeneity in the rankings and has already tweaked its software to try to address it. So far, he said, the data suggest that injecting more diversity into users’ ranked timelines might actually be good for engagement. If that’s true, it could be good news for both the company and its users. Then again, recent history is rife with examples of tech companies’ interests not aligning with society’s quite so neatly.
* * *
At a time when Twitter is still clawing desperately for traction—it has launched multiple live-streaming products and overhauled its harassment policies in a bid to budge its stagnant user growth and clean up the more noxious and abusive discourse in its commons—giving the algorithm more control over users’ feeds would be a logical next move. It would be risky, sure. But for Twitter, at this juncture, the risks of inaction may be greater.
Yet if ever-greater personalization is the answer to Twitter’s business woes, it’s unlikely to be the answer to the woes of a media ecosystem in which all news has become “fake news” to someone. Presenting people with contrary viewpoints simply isn’t a recipe for massive gains in engagement, for reasons that may run too deep in human nature to change.
On the other hand, the main alternatives to Twitter as a news source—Fox News, CNN, Facebook, et al.—all have their own crises of credibility and perception, driven in part by the perverse incentives of their respective business and audience models. If Twitter is right that its users value at least some diversity of viewpoints in their feeds—and if it’s serious about being a place people come for information rather than just entertainment or endorphins—we’ll be better off with a more automated Twitter than we would be with no Twitter at all.