How many hours should you stream? The gains stop at eight

Six to eight hours in a sitting, and after that the numbers stop moving. Across 1,973 live Grand Theft Auto streams, measured ten times between 04:06:03 and 04:16:03 UTC this morning, a stream that had been live for six to eight hours was clearing ten concurrent viewers 45.4% of the time. Under two hours: 27.1%. Past twenty-four hours: 36.3%, which is back below where a stream sitting at the four-hour mark already was.

That is the short answer, and it is roughly twice the two-to-four hours the advice pages recommend. The longer answer is the one that matters more, and it is uncomfortable: across the hundred biggest GTA creators on the planet, the number of hours someone streamed in the last month has no relationship at all with how many people watch them. Rank correlation: −0.11.

What the guides say

The pages that rank for this question agree with each other and cite nothing. Streamerfacts tells small streamers to “stream between 2 and 4 days a week” for “2 – 4 hours per day”, reasoning that shorter streams mean viewers “miss a large part of the stream”. StreamScheme lands in the same place. Neither is unreasonable. Neither has looked at a dataset.

We sample every live GTA stream on Kick, Twitch and YouTube once a minute, and every live stream carries how long it has been running. So this is answerable.

The session-length curve

Hours live	Streams	Clears 10 viewers	Range across censuses	Median viewers
Under 2	571	27.1%	26.4–27.8%	3
2–4	561	35.6%	34.6–36.5%	5
4–6	366	41.6%	41.0–42.2%	6
6–8	196	45.4%	43.1–47.2%	8
8–12	176	44.6%	43.1–46.3%	8
12–24	56	38.9%	35.7–41.8%	5
24+	47	36.3%	34.0–37.0%	2

Ten complete censuses, 04:06:03–04:16:03 UTC, 7 September 2026. Stream counts are the mean per census; the board and its tail reconciled exactly to summary.json in all ten, with no duplicate channels.

The shape is a hill. It climbs steeply for the first six hours, flattens across six to twelve, and then falls away. The forty-seven streams that had been running more than a day had a median of two viewers — the lowest figure in the table, below every band including the under-two-hours one they all started in.

The obvious objection, which is right

A stream that has been live eight hours at 04:10 UTC started at 20:10 UTC. A stream one hour in started at 03:10. So the column labelled “hours live” is also a column of start times, and 20:10 UTC is prime time across the Americas while 03:10 is nobody’s. We have measured that effect before and it is large — mean viewers per channel nearly triples between the 09:00 UTC trough and the 21:00 peak.

We cannot separate the two from a single hour of censuses, and this post is not going to pretend otherwise. What the clock does not explain is the far end of the table. A stream twenty-six hours in started at roughly the same clock hour as one two hours in, one day earlier, and it is doing worse than either. Whatever the six-to-eight-hour peak is made of, the collapse after twelve hours is not a start-time artefact.

There is a second reading of the same shape, and it is the deflating one: sessions do not run long because they are doing well. They run long because subathons and 24/7 marathons are a format, and that format is popular with people whose audience is small enough that the hours are affordable.

Thirty days, a hundred creators, and nothing

The census answers “how long should tonight be”. It cannot answer “how often should I go live”, because it only sees the session in front of it. For that there is the creator board: the hundred largest GTA creators across all three platforms, with total hours streamed and average concurrent audience over a rolling thirty days.

Ninety-nine of them have recorded airtime. Their hours range from 0.1 to 359.4 across the month. Their median is 8.4 hours — for the month, not the week. And hours predict the audience not at all:

30-day hours	Creators	Median hours	Median avg. viewers	Median hours watched
Under 2.6	24	0.8	10,402	11,412
2.6–8.4	25	4.2	11,546	61,331
8.4–28.9	25	13.5	13,987	225,335
Over 28.9	25	85.6	8,431	759,973

Read the last two columns against each other, because that is the whole finding. Going from four hours a month to eighty-six multiplies hours watched by twelve. It multiplies average audience by 0.73.

The two extremes make it concrete. 3mr, an Arabic-language Kick channel, streamed 359.4 hours of GTA in thirty days — about twelve hours a day, every day — and averaged 13,132 concurrent viewers. sxb streamed 19.4 hours, 5.4% as much, and averaged 128,802. And IShowSpeed appears on this board on 0.2 hours of GTA — twelve minutes — at an average of 140,134, the largest in the category.

Rank the hundred by hours watched and rank them by average viewers, and the two top tens share three names.

The correlation between hours and average audience is −0.11 across all ninety-nine, and it gets more negative, not less, when you drop the cameo appearances: among the thirty creators who streamed at least twenty hours in the month, it is −0.55. Hours and audience move in opposite directions at the top of this category.

So what is a creator supposed to do with that

Not “stream less”. The arrow almost certainly runs the other way — people with enormous audiences can afford to go live twelve minutes a month, and people building one cannot. Nothing here says quitting the schedule would have made 3mr bigger.

What it does say is that the specific promise inside the standard advice is not in the data. “Stream more and your concurrent numbers will rise” is the reason people are told to grind, and at the top of GTA the grind is uncorrelated with the number. What the hours reliably buy is hours watched — that correlation is +0.93, close to mechanical, because watch time is roughly audience times time. If you are selling ad impressions, sponsor deliverables or a subathon, that is the number that pays and the hours are the way to get it. If what you want is a bigger room, the hours are not the lever.

Which is worth knowing before you commit a year of evenings to it.

What this does not show

Four things, plainly:

• The board is the top hundred by average audience, so the range of audience sizes is cut off at the bottom. Truncating a variable flattens correlations toward zero, which means −0.11 is a weaker piece of evidence than it looks — and also that this is a finding about large creators, not about someone at nine viewers deciding whether to stream on Tuesday.
• hours live is elapsed time, not final session length. A stream two hours in may run another ten. Long sessions are also over-represented in any snapshot, because a twenty-four-hour stream is visible in twenty-four hourly censuses and a two-hour one in two.
• One clock hour. The census is 04:0x–04:1x UTC, which under-samples European and North American evenings. The thirty-day board does not have that problem; the session-length table does.
• We count concurrent viewers, not people, and not attention. A viewer parked on a marathon overnight counts the same as one watching.

Check it yourself

Everything above comes from two public files. The live census is data.luciastream.com/live/streams.json plus live/streams-tail.json — the board is the top 400 and the tail is the rest, and you need both — sampled ten times between 04:06:03 and 04:16:03 UTC on 7 September 2026. The thirty-day creator figures are data.luciastream.com/boards/top-creators-30d.json, read at 03:56:41 UTC the same morning. Method is at /methodology/, the live numbers are at /live/, and the creator board is at /leaderboard/.

Permalink: https://luciastream.com/blog/how-many-hours-should-you-stream/. If you use these figures, the sampling window is the part that matters — a GTA audience number without a timestamp on it is not a number.

One note on that board for anyone reading it directly: the exporter writes the same thirty-day aggregate to the 7d, 30d and all paths, so all three URLs return identical rows. The window in the filename is not doing anything. Treat every figure on it as thirty days.