Algorithmic Timelines

There’s often a backlash when an internet service changes something user-facing. Occasionally these lead to companies backpedalling (eg. Patreon’s changes to their fee structures), being quickly supplanted by a competitor (eg. Digg’s redesign giving Reddit the “what’s hot on the internet” crown), or sinking into irrelevancy (eg. YikYak’s move away from anonymity). But most outrages die out quickly, and soon become a non-issue (eg. stars into hearts, round avatars, or Facebook Messenger becoming its own app).

But one recent change that resists a rollback/replace/acclimatise categorisation is the introduction of algorithmic timelines. When you opened Twitter or Instagram in the ancient days of 2016, your feed was in chronological order: newer posts first, and older posts towards the bottom[1]. But no longer: an algorithm decides the order now. It seems like everybody hates it, but the feature remains. Why? Why would a company keep a feature that their users continue to hate? The general consensus seems to be that it’s about advertising revenue, but I think that’s a second-order effect[2]. My own take is it’s really about Dunbar’s number and stickiness.

Dunbar’s Number

Robin Dunbar proposed that social groups have an innate natural limit, based on our brain’s ability to keep track of who’s in a group. This limit – known as Dunbar’s number – is often stated as a firm 150, but was originally a little softer; somewhere between 100 and 250 people. It’s generally seen as an inflection point for organisations, because once a company grows beyond Dunbar’s number its employees can’t know everyone who works there any more. Companies generally solve this by forming divisions or teams – smaller groups where you can know everybody – and having defined channels of communication between those groups. Social groups might split into separate organisations too – if the Toronto Bridge Club grows a lot, then it might split into the East Toronto Bridge Club & West Toronto Bridge Club.

There’s a similar effect on a growing social network. Users follow a lot of people at first, because they want more connections and more content. But it eventually becomes too much to keep up with, and users start being way more selective about who they follow. The exact point varies[3], but for me it’s about 500 people on Twitter and just under 200 on Instagram. And it’s not just me: I did a little data science to demonstrate this threshold, and wrote some code that checks the people I follow on Twitter & stores how many people they follow. Here’s the results:

[[[[[[[results here]]]]]]]]]

Let’s play “Fantasy Product Manager”

Imagine you worked somewhere that had “made it” – tens of millions of active users, part of the popular consciousness, and a history of quarter-over-quarter growth. Wouldn’t those charts keep you up at night? You’ve done two really hard things: gained a critical mass of users, and kept them interested over time. But now your growth is slowing and you’re facing that wall. There’s not many new users to acquire – anyone who wants an account already has one – and your existing users aren’t following new people any more. If they did, they’d start losing track of people or feel like they’re not keeping up. So what do you do?

You know that posts are not all created equal. Some are better — funnier, cooler, smarter, sexier, more urgent, more interesting, more emotive — than others. What if you showed people the best posts first? They won’t worry about missing things if they trust your product to always show them the best content. Nobody feels anxious about missing important things in a newspaper – that’s what the front page & the headlines are for. The underlying fear hidden in “not keeping up” is “I’m worried I’ll miss something important”. But now you’re promising to always show the best stuff first! Problem solved! 🚀

There’s two big flaws in this plan, though. Algorithms can’t perfectly rank posts from best to worst, and it’s an open question whether they can even reach “good enough”. For all the spookiness of ads recommendations there’s still thousands of stories of life events missed and complete mishits. But even if you can overcome that problem, users still lose the feeling of understanding your service and being in control. “Newest stuff at the top, it gets older as you scroll” is easy to grok. You know you’re caught up when you see things you’d seen before. But Instagram now? There’s literally no way to tell, because there’s no stable ordering.

This is a fundamental change in the product. It’s no longer a “Keep up with content you care about” kind of thing[4]; it’s a “Here’s a slice of stuff you might like” kind of thing[5]. Twitter’s been more cautious about this than Instagram (pitching it as “In case you missed it”), but I’d bet this is an intentional change. Instagram’s move to let you subscribe to hashtags feels like another step towards this goal: nobody can view every single post for a hashtag, but you can see the new hotness every day.

Stickiness

We’ve established that algorithmic feeds might help a service overcome Dunbar’s number, if they’re willing to shift their operating model to one where people skim for their best stuff. Let’s talk about stickiness, the second potential outcome. You can measure stickiness in a variety of ways:

• How long do users spend in your app?
• How often do they come back?
• How much time do they spend once they come back? How is this affected by how long they were away? Everybody’s closed Facebook only to immediately reopen it.

Let’s explore that last scenario some more. If your product has a chronological feed, there’s nothing new for your user. Maybe, if you’re really lucky, you’ll have a new post or two. But even if you do, the user’s going to catch themselves & feel like a dummy. Closing an app only to reopen it! How stupid can you get?[6]

Maybe your snazzy algorithm can fix this. You’ve already got a quality rating for the user’s posts, and you know which ones the user’s already seen. So why not show them something less good, but new to them? They’ll be more likely to stick around, they won’t feel silly, and you just re-engaged a user. More stickiness, and you’re giving the user the quality fresh content they were looking for. Job done, and everybody’s happy.

And yes, these changes lead to more ad revenues. More users means more ad inventory[7]; more engagement means higher ad prices. I reckon that’s why these features have endured despite widespread user distaste: they work, or at least make the key metrics move in the right direction. Is it a short-term or long-term movement? There’s no way to know for sure yet. Users may not register the cognitive dissonance of being more engaged with an app they claim to dislike, but they may eventually make a change. Maybe they’ll switch to a competitor, or find something entirely different to fill their time. Maybe their previous love will fade to a slightly-positive indifference. No matter which, the company may not care: huge companies always overlook small user cohorts. They don’t think about individual users any more – it’s the aggregate numbers that matter.

What happens next

I don’t have a prediction for how algorithmic timelines will play out. Twitter is far from an ideal company, but most of their changes are positively received in retrospect – plus their algorithmicness has been contained in a small section of their product. It’s been 7 years[8] since Facebook showed everything posted – 60% of users[9] never even used that version – so they’re obviously going to stick with it. Instagram, though, is harder to predict: it seems like people are more unhappy with them, but their changes also seem more like a deliberate paradigmatic shift. I don’t know which metrics they’ve chosen or how they’ve changed since the introduction, so it’s hard to predict whether they’ll stick with it or reintroduce an option for a pure feed. I suspect they’ll stick with it, mostly because their product doesn’t let users customise it much.

Even though most services aren’t hitting these growth walls, I think algorithmic feeds will continue to proliferate (despite Ian Malcolm’s warning). Machine learning is the new hotness and becomes ever-easier to integrate. But please, I beg you: if you must follow the herd, do it in a light touch way. Separate your algorithmic content visually into a separate section, or intersperse it with your chronological feed. Let users flip between both views, and don’t be upset if they prefer the former. With a light touch integration, you can have your cake and eat it too.