Do subathons work? The hours rise faster than the viewers

A German GTA roleplayer who streams as Dayi went live at 15:59 UTC on 1 September and was on air for 106 of the next 110 hours, in three blocks, against the three-to-seven-hour evenings he had been doing the week before.

His average concurrent audience across those 106 hours was 421. Across his nine recorded broadcasts before it, 423.

That is one person and one game, so hold it loosely. But it is the shape of what we found this morning when we went looking for every GTA subathon we could measure. Across the five where a creator can be compared against themselves, the median one gained 11% more concurrent viewers while running twice the hours per session. The two that did much better than that were both playing NoPixel V, and the thirteen channels in our control group who were also playing NoPixel V — and running no subathon — gained a median of 85%.

What a subathon actually is, measured

Page one for this question is six explainers — Streamlabs, StreamScheme, StreamYard, Dot Esports, GankNow, Overlays.uno — and they agree on the definition. A subathon is a stream with a countdown timer that every new subscription extends, so the broadcast runs until the subs stop coming. They agree on the benefit too: it compounds watch time, it surfaces you in recommendations, it is “an excellent way for smaller streamers to grow their audience”. Between them they publish no number at all about whether that last part is true.

The first thing our data says is that the definition is wrong, or at least that almost nobody plays it that way.

Across twelve complete censuses of every live GTA stream on Kick, Twitch and YouTube, taken a minute apart between 04:11:33 and 04:23:33 UTC today, 21 channels advertised a subathon in the stream title. The same 21 appeared in all twelve. They are 1.11% of the 1,894 channels live and hold 3.99% of the 269,481 viewers watching, though that share is misleading on its own: the largest three hold 88% of it between them.

Twelve of the 21 print a day number, and the numbers run from day 1 to day 64, median 10. If a subathon were the continuous timer-driven marathon the guides describe, a channel on day 14 would have been live for a fortnight. None of them has. The median subathon stream in our census had been continuously live for 9.1 hours — a long evening — and exactly one of the 21 was past a day, at 40.0 hours, on day 64.

So in practice a GTA subathon is a run of ordinary daily sessions with a counter in the title and a sub goal attached, not a single unbroken broadcast. That matters for anyone reading the advice, because the effort being described and the effort actually being spent are different things. It also means this is not the same population as the channels in our 24/7 rebroadcast piece from Sunday, which were continuously live for a week or more and held 93 viewers between them.

One aside for anyone matching titles themselves: SUBtember is not a subathon. Twitch’s September discount promotion puts the word in dozens of GTA titles this month, and a naive keyword match returns 64 channels instead of 21. We threw those out. We also separate six streams advertising a “marathon” or a “24/7 until day N” challenge, which are endurance streams without a sub timer; the largest, on Kick, is discussed below.

The test: each creator against themselves, and against their peers

The trouble with asking whether subathons work is that the people who run them are not a random sample. A creator with a subathon is a creator with a plan, an overlay and probably a community, and comparing them against the category median measures that instead.

So we did it the other way. Our per-creator data carries the last twelve broadcasts for every channel we track, each with its own duration and its own airtime-weighted average concurrent viewership. For any creator whose subathon began inside that window, we can compare them to themselves: the sessions before against the sessions during.

That still leaves the week itself. GTA viewership has been unusually high since NoPixel V opened on 8 September, so a creator who started a subathon on the 8th would have risen anyway. For each subathon we therefore built a control from the same morning’s census: Twitch GTA channels between 62 and 1,966 concurrent viewers, no subathon or marathon in the title, with at least two sessions on each side of that creator’s own start date. Thirty-one to forty-eight channels qualified per cut-off, and we split them by whether they were on NoPixel V before taking the median.

Five creators had a clean before-and-after boundary. Here is all five.

creator	began	on NoPixel V	sessions before / during	viewers before	during	change	like-for-like peers
NovaaBerries	9 Sep	yes	8 / 4	41	138	×3.39	×1.78 (n=13)
Chatterbox	8 Sep	yes	7 / 5	1,475	4,234	×2.87	×1.85 (n=13)
SupaDupaTroy	7 Sep	no	5 / 7	47	52	×1.11	×0.99 (n=34)
Dayi	1 Sep	no	9 / 3	423	421	×1.00	×0.86 (n=20)
DRZL	1 Sep	no	5 / 7	158	153	×0.97	×0.86 (n=20)

All five are on Twitch, and the peer column compares each to the control channels that were doing the same thing that week — NoPixel V or not — which turns out to matter more than anything else in the table. We come back to that below.

Every one of the five finished a little ahead of their peers, and three of them did it by moving their own audience −3% to +11% while the peer group drifted down. First, look at what that cost them.

DRZL, who plays a police officer called Gary Snackwell on a server named Starlight, ran seven subathon sessions between 1 and 13 September. Five of them were over 34 hours long. He was on air for 194.9 hours in twelve days, against evenings of six and a half to nine and a half hours before he started. His average audience went from 158 to 153.

Dayi’s numbers are the ones that stopped me. Across all twelve of his recorded broadcasts — the nine ordinary evenings and the three subathon blocks, the shortest 3.7 hours and the longest 44.4 — his average concurrent audience never once left the range 417 to 431. Whatever he did with the clock, the same four hundred people watched.

The Kick 24/7 challenge we set aside earlier tells the same story from a different direction. rampagejackson has been running “LIVE 24/7 UNTIL… DAY n” since late August, in sessions of 24 to 27 hours. His per-session average across them runs 5,102 to 6,306. It was 5,987 on his last ordinary evening stream before the challenge began.

The two that tripled were both on NoPixel V

Chatterbox nearly tripled. NovaaBerries more than tripled. Both are real, and the explanation is almost certainly not the subathon.

Chatterbox’s subathon day 1 was 8 September, which is the day NoPixel V opened. Her own title that night read “NoPixel V FIRST DAY! | !SUBATHON DAY 1”. She did not run a subathon and get bigger; she got an invitation to the largest event in GTA roleplay in years and ran a subathon on top of it.

Split the control group by that and the confound stops being a hypothetical. Of the 48 comparable Twitch GTA channels with sessions either side of 8 September, 13 were playing NoPixel V and 35 were not — and not one of the 48 was running a subathon.

control group, 8 Sep cut	channels	median change	best
Twitch GTA, playing NoPixel V	13	×1.85	×7.94
Twitch GTA, not playing NoPixel V	35	×1.00	×2.23

Being on the new server was worth a median of 85% that week without any sub timer at all. Not being on it was worth nothing: ×1.00, the flattest possible result. The split holds at every cut-off we tried between 1 and 9 September, with the NoPixel median between ×1.78 and ×2.06 and the rest between ×0.86 and ×1.00.

Drop Chatterbox and NovaaBerries into the group they actually belong to and they stop being outliers. Each would rank third of fourteen — ahead of the NoPixel median, behind two channels that gained ×5.87 and ×7.94 with no subathon anywhere in sight. That is a good week inside a very good week, and with two observations we have no business calling it the timer.

Meanwhile the three subathons outside NoPixel V beat peer groups that were themselves flat or falling, by 11 to 14 percentage points, on individual gains of −3% to +11%. Whether a margin that size survives a bigger sample is exactly the question we cannot answer with three people.

NovaaBerries’ number is thinner still on its own terms. Her ×3.39 rests on four sessions, one of which — 11 September, peaking at 1,312 against pre-subathon peaks of 42 to 100 — carries most of it. That is the signature of a raid or a front-page placement rather than a compounding timer, and we cannot see which.

What we cannot measure, and it is the important part

A subathon is not primarily an audience tactic. It is a subscription drive. The timer is attached to subs, the goal is stated in subs, and the money is the point.

We do not measure subscriptions. We count concurrent viewers, once a minute, from each platform’s public API. Nothing in this post is evidence about whether a subathon earns anyone a living, and it is entirely possible for a subathon to be the best financial night of a creator’s year while the viewer number sits flat — which, for a channel converting an existing audience rather than reaching a new one, is roughly what you would expect to see.

We checked whether we could say something about follower growth instead and decided we could not. The follower series behind three of these five channels is unusable: two are frozen at the same integer for a fortnight, and a third jumps by exactly 5,000 in a single day. We would rather print that than a number built on it.

The rest of the limitations, briefly. Five creators is a small number and they are all on Twitch, because the Kick and YouTube subathons in our census did not have enough sessions on both sides of their start date. Our session history stops at twelve broadcasts, so a subathon older than that window has no “before” for us to find, which biases this sample toward recent starts — and recent, this month, means NoPixel V. The census is a fifteen-minute window near the daily trough.

Both of the splits in this piece are read off stream titles, which is a real weakness and cuts in a known direction. A subathon not advertised in the title is invisible to us. More importantly, we sort the control group by whether “NoPixel” appears in the title, and when we checked this two days ago 76.6% of live GTA roleplay streams named no server at all. Some NoPixel V channels are therefore sitting in our “not playing NoPixel V” column, which would pull that group’s ×1.00 up, not down — so the gap between the two control groups is if anything understated.

And nothing here is causal in either direction: a creator who calls a subathon during a quiet patch and a creator who calls one during a purple patch will both show up in this table, and we cannot tell them apart.

So should you run one?

Not for the viewers, on this evidence. The honest summary of the five is that hours on air roughly doubled at the median, the audience barely moved, and what edge there was over their peers — eleven to fourteen points, on three people — is the kind of number that needs a much bigger sample before anyone builds a schedule on it. It is also the third time our data has pointed this way, after session length and total hours streamed, both of which found time on air to be a far weaker lever than the advice industry treats it as.

If you are running one to convert an audience you already have into subscribers, none of this touches that, and it may well work. If you are running one because six explainers told you that uptime compounds into reach, the strongest thing we can say is that the two channels here that tripled did it in the week they joined the biggest roleplay server launch of the decade, and that being on that server was worth ×1.85 to channels running no subathon at all. The two who spent 106 and 195 hours on air outside it moved their audience by 0% and −3%.

Pick the server, not the clock.

Check us

Every figure above comes from twelve complete censuses of every live GTA stream on Kick, Twitch and YouTube between 04:11:33 and 04:23:33 UTC on 15 September 2026, taken from data.luciastream.com, plus the per-creator session histories published alongside them. In all twelve the top-400 board and the tail file carried matching timestamps, the per-stream counts reconciled exactly to the published totals, no channel appeared twice, and no platform reported degraded coverage. The mean across the twelve was 269,481 viewers on 1,894 channels, individual samples ranging from 253,210 to 283,636, with the board cut-off drifting between 43 and 49 viewers.

Session-level figures are each platform’s own reported start time and duration, taken at face value, and viewer averages are airtime-weighted within each session.

The live board is at /live/, the creator leaderboard at /leaderboard/, collection is described on our methodology page, and the permalink for citation is https://luciastream.com/blog/do-subathons-work/.