How we measure

The window

We read each channel's completed streams from the last 14 days. A stream counts only once it has ended, so partial live sessions never skew a figure. Within a stream we focus on the first 120 observed minutes after go-live, the opening window where the audience assembles.

Viewer ramp-up

The ramp-up curve plots the channel's viewer count from its real go-live (the Twitch start time), one point per reading. We read most channels every few minutes, and more often in the opening window, so the curve is a series of readings joined up, not a continuous recording. It shows how the room fills over the opening window, not just how big it eventually gets.

Time to half crowd

First we take a sustained level: the raid-excluded 90th percentile of the viewer counts we read during the stream, so a brief raid spike never sets the bar. Time to half crowd is the minutes from the real go-live to the first observed minute the audience reaches 50% of that sustained level, averaged across the streams where both arrival points were measured. It answers how quickly half the room is in.

Audience arrived

Audience arrived uses the same sustained level and is the first minute the curve holds at least 90% of it for three consecutive readings. Before that minute the room is still filling. Because both points measure against the same level, half the room is always in before (or at) the moment everyone is in.

Late-arrival gap

For each stream we take the opening-window peak, the biggest raid-excluded crowd inside the opening window. Then, for every reading, we work out how far the crowd was below that peak and multiply it by the minutes until the next reading, and we add those products up. That total is the stream's shortfall in viewer-minutes. Because each reading is weighted by the time it covers, the figure comes out the same whether that stream happened to be read every two minutes or every five. Averaged over the measured streams, that shortfall is the late-arrival gap. It is the audience that missed the opening while the room was still filling, and the moment a Ping helps fans show up at minute one.

The opening window runs up to the first 120 observed minutes; shorter streams use their observed span, shown per stream. It is long enough to contain most channels' ramps in our observations, and it is the window a go-live Ping targets.

Which channels we compare against

The comparison set is the channels fanping actively observes, a large roster we build by discovering channels through Twitch's public directory. We group them into size lanes by the crowd a channel typically sustains (the median of its streams' sustained levels, not one-off peaks), and you are only ever ranked within your own lane. This set is not all of Twitch: it is the channels fanping has seen, so every "of N channels your size" figure names that observed population, never the whole platform.

Comparisons and lanes

Where a channel is compared, it is compared with similar channels: a size lane of channels with a comparable audience size, over the same 14-day window of completed streams.

A comparison is shown only when its lane holds at least 8 similar channels. Below that we suppress it ("Not enough similar streamers yet"), so a standing is never drawn from too few peers. Channels whose streams span several games get a looser comparison, and we say so.

What we adjust for

To keep every figure honest we leave out the moments and streams that would distort it. Here is exactly what we set aside.

• When another channel raids in, the viewer spike it brings is set aside for a 10-minute window (plus the 2 minutes before it lands), so a raid never sets the opening peak, the session-max marker, or which size lane a channel sits in.
• Streams whose go-live we did not catch (we started watching late, or the stream was already warm) carry no opening landmarks and are left out of the opening-window averages, and we name them in the per-stream breakdown.
• A stream that is still live is never counted in the 14-day averages; it joins only once it has ended.
• Arrival times longer than four hours are dropped from the arrival averages and rankings, because a marathon stream's daily rhythm is not an opening-ramp figure.
• Streams whose opening spike sits far above the level they go on to hold are left out of the late-arrival gap average and named in the breakdown, since that usually means an undetected raid or a front-page burst rather than a real early crowd.
• Channels that look like bots, and streams shorter than 15 observed minutes, are left out of comparisons.
• We read viewer counts about every five minutes, so a brief peak between two readings can go unseen.

Removing a channel

fanping creators can turn this off in Settings. Everyone else can opt out with a verified Twitch check. See the start page for the common questions.