Going viral
The surprising road to millions of YouTube views
Back in summer 2016, I gave two talks on a similar topic in the space of a week. One was at London’s Royal Institution, the other at Google. Both talks were posted on YouTube afterwards, on two channels with similar numbers of subscribers. Then something strange happened. In the next a year or so, the Royal Institution talk racked up over a million views. In contrast, the Google talk was only viewed by around 10,000 people.
I was curious about what had happened, so I asked the Royal Institution if I could analyse the data. And that’s when I got another surprise, which would challenge my preconceptions about online virality.
The data showed that after the Royal Institution video had been posted, it hadn’t attracted that much attention for the first few months, receiving around a hundred views per day. Then the pattern changed. In a matter of days, it got more views that it had received in that entire first year. The numbers would continue to accumulate at a rate of a thousand or so per day; the total now stands at just over 3.5 million.
Had word-of-mouth recommendations somehow driven vast numbers to my talk? Maybe the video had gained traction in some corner of social media, going viral as it was reshared from user-to-user? The reality turned out to be very different. YouTube had simply featured the video on their homepage. As more and more people watched, the algorithm put it on the ‘suggested video’ lists that accompany other popular videos. Of the people who viewed the talk, almost 90% came directly from the homepage or one of these lists.
Super-superspreading
During a disease outbreak like influenza or COVID, infections spread from one person to another in a sequence of transmission events. Some online trends spread in a similar manner, with each participant exposing a handful of ‘susceptible’ others to the outbreak, who in turn expose additional people. Take the ‘neknomination’ trend in 2014, where people filmed themselves downing a drink and nominated a couple of friends to do the same. Given the similarities, I’d even analysed the game in its early stages for BBC Radio.
But not all online contagion spreads like a nominated-based game. One 2016 study of popular content on Twitter found that, of the 0.03% of Twitter posts that ended up getting over 100 shares, very few spread through a long sequence of transmission links from user-to-user-to-user. Insead, their success typically came down to a single ‘broadcast’, with a large number of users picking up the content from a single source. Rather than the dozens of susceptible individuals that might be involved in a biological superspreading event, online outbreaks can involve super-superspreading events, with thousands or millions potentially exposed at once.
My Royal Institution talk had been a textbook broadcast event, with one source (the YouTube recommendation algorithm) responsible for almost all of the views. It wasn’t a story of virality so much as amplification from big organisations: first the Royal Institution channel to generate a few thousand views, then the YouTube algorithm bringing in the millions.
Sparking contagion
If online content doesn’t spread easily from person-to-person in a long chain of transmission, we therefore need to rethink how to get ideas and content out into the world. In 2007, Duncan Watts – who also led the above study of Twitter data – co-authored an analysis of several marketing campaigns. The team found that the reproduction number, R, of each campaign (i.e. the average number of new people each person shared the message onto) was generally very low. In fact, it was so low that transmission could not persist on its own for long.
One campaign related to Hurricane Katrina donations: for every 100 people contacted, the message was passed onto 77 others on average (i.e. R=0.77). In the weakest campaign, relating to detergent, the message was apparently only passed on to 4 other people (i.e. R=0.04). The team also analysed campaigns by the advertising agency J. Walter Thompson, who’d bought all the ad space on the Huffington Post for a week. Analysis of e-mail sharing found that the seven ads they’d displayed had a reproduction number ranging from 0.5 to 0.9.
These results suggest that content is unlikely to spread widely on its own after a single tweet or e-mail. Instead, it will typically need many introductions and amplifications, each of which may subsequently spark a small outbreak that generates some additional sharing.
One of the co-authors of the marketing analysis was Jonah Peretti, who had recently founded Buzzfeed. When I later wrote the The Rules of Contagion, I covered how websites like Buzzfeed and Huffington Post used these epidemiological ideas to become leaders in viral content during the 2010s. When Tom Chivers (ex-Buzzfeed) later reviewed the book for The Times, he gave an insider insight into just how direct this link was:
When I worked at Buzzfeed, the website’s back end provided authors with a simple piece of information, “social lift”, based on the ratio of clicks via Buzzfeed’s website to clicks from social sharing. A social lift of two means that every person clicking on a story through the website leads, on average, to one person clicking on it from elsewhere. It occurred to me, as I was reading, that this was precisely analogous to R: a social lift of two equals an R of one. The “viral” analogy was more explicit than I realised.
Finding the right explanation
When people talk about online virality, I’ve noticed they usually focus on the content itself. Make something interesting or entertaining, goes the reasoning, and that will make it spread. Certain characteristics of content – like novelty, usefulness, or a tendency to trigger strong emotions – can boost its shareability, but success ultimately depends on how people are exposed to that content. And in most cases, that means having a broadcast event somewhere.
So next time you see something become popular online, why not try and investigate where the popularity really came from. Because, as I’ve learned, it might not lead to the explanation you expect.
This is exactly how our TED talks work. It has very little to do with the quality of them. I graphed my views over time, and it was a delta function once TED put it on the front page. The longer it stayed up there, the more views it gets, by ~500k per day, plus a long dying off period. All of the TED fellows talks that were posted had the *exact* same trajectory based on when theirs was put on the front page (usually for 24 hours) vs when the link was active. Also another friend that year had her regular TED talk featured for 3 days, and hers correspondingly had another 500k+ views. It's **entirely** driven by what the site chooses to promote. Sure, there is a link to quality, but it's most strongly determined by the marketing from these sites themselves. Same with amazon links etc. You have to get lucky or have it promoted by the marketplace.
I think this is really interesting in the context of Mastodon and Threads. I’m on both. I think the kind of promotion discussed is only possible on Threads. I like both platforms but use them very differently.