Language not available Language not available Language not available

This post is not available in your language. Here are some other options:

Who Hosts Who? Who Hosts You?

Feb 17 2015

When we launched Host Mode in late July 2014, we weren’t sure how people would use the feature. The ability for broadcasters to choose to host another live channel when they weren’t broadcasting has lots of potential uses:

Now that Host Mode has been in the wild for some time, we wanted to share some insights about how it’s been used and the impact it’s having on the Twitch community. In particular, we wanted to explore one specific question: What kinds of channels are using Host Mode and who are they hosting?

Which channels use host mode?

One of the primary differences between channels is how many viewers they have. On one end of the scale we have mega channels with tens of thousands of concurrent viewers. On the other end, we have emerging broadcasters that are just getting started. Because the scales here are so extreme, we’re going to use a log scale to bin broadcasters; this means the values on the x axis in these graphs are a measure of the order of magnitude of the size of a channel’s audience.

So which kinds of channels are using Host Mode? Is it used more by emerging channels or big, established channels? Popularity of channels (or anything, really!) is often modeled as a power law like this one:

Since anyone can broadcast (and lots of people only broadcast once or twice to a very small audience), this is a reasonable approximation of how viewers are distributed on Twitch; most channels have a small audience, and a few channels (the right side of the distribution) have a very large audience.

Clearly Host Mode usage isn’t random; if it were, this distribution should line up more or less identically with the overall popularity distribution. Instead, we see that the majority of uses are centered around channels in the middle of the popularity spectrum.

Who are they hosting?

Our next big question is who are they hosting? We had three basic hypotheses here:

**Lets look at the same distribution of usage, but color it based on the relative popularity of the host and target of hosting. If they’re about the same popularity, we’ll call it **peering; if the host is much more popular than the target, it’s spreading; and if the target is much more popular than the host, then it’s supporting.

There’s lots going on here! First, we can see that peer-type hosting is the most popular overall by far, followed by supporting, and then spreading. Remember that this is counting discrete hosting instances; each time a broadcaster starts hosting, it counts once in this graph.

We have to be a little careful when comparing different popularity buckets. At low popularity, there are no channels less popular for you to host, so the lack of spreading isn’t meaningful. And at high popularity, there are few people more popular than you, so supporting will necessarily disappear. But in the middle, when there are plenty of channels more or less popular, we can see that people still prefer hosting channels that are a similar level of popularity to them over the other potential options.

What is the viewer perspective?

So far, we’ve been focused on the decisions that broadcasters make about whether to host and who to host. But what about the viewer perspective? The more popular the channel, the more people will see Host Mode, so viewers experience the effect of hosting much more when really popular channels do it.

Lets take the above graph, but rescale it by how many people actually saw the channel when it was in Host Mode. So if a channel with 1,000 concurrent viewers turns on Host Mode we’ll count that 1,000 times, instead of just once, as in the previous graph. If we sum up these views, we can see the overall split between each of the three behaviors both in terms of how often broadcasters do each one and how often viewers see them.

Starts Views Support ■ 28.3% 6.1% Peer ■ 62.0% 74.9% Spread ■ 9.7%


With this perspective, we see supporting (emerging channels hosting more popular channels) doesn’t drive much traffic, even though it happens pretty frequently. This makes sense, since that behavior is most common among emerging channels. Peering gets even more popular, and spreading (which is more common among higher popularity channels) drives nearly 20% of Host Mode views.

This gives us a clearer picture of how Host Mode is being used by the broadcaster community. Our focus at Twitch is supporting broadcasters and enabling channels of all sizes to grow. Both of the behaviors we’ve seen point to this as well. When given the choice, broadcasters will use Host Mode to spread their audience to channels of a similar size or smaller, rather than to more popular channels that they are more likely to know about. We think that’s awesome, and we’re glad to see the community showing this kind of support for their fellow broadcasters!

Still with us? Want to help us do work like this every day? We’re hiring researchers and data scientists to join our Science team! Check out our job posting and learn more about our team. Shoot us an email if you’d like to learn more about the role or apply!

In other news