We shut it down.
The measurements stay.

Channel Autopilot · started 2026-09-18 · ended 2026-09-28

It was a service where an agent took a topic, made a Short, and uploaded it to your YouTube channel every day. It ended with zero paying customers. We stopped selling it, but there was no reason to bury six months of numbers with it, so they are here.

What we measured

Channels run fully unattended	7
Videos published	498
Total views	176,580
Median views per video	90
Videos above 10,000 views	0
Best single video	4,015
Like rate	1.72%
Cost — still-image video	$0.084 / video
Cost — AI video	$2.08 / video
Paying customers	0

Sources: full census via the YouTube Data API, and Google Cloud invoices. Measured, not estimated.

The wall was not production. It was channel size.

A benchmark channel had 30,900 subscribers and 200,000 views per video. Our best channel had 165 subscribers and 1,149. Subscribers differed by 187×; views by 177× — almost exactly proportional. Meanwhile our subscriber conversion was normal (1.7–4.4 per thousand views; benchmark 2.46). So the videos were not the problem. Reach is set by channel size, and that is the one thing we could not hand anyone. Driving production cost down to $0.08 a video does not move it.

What we learned

“Nobody does X” and “nobody can do X” produce identical data. When a number is zero, check whether a path existed before blaming demand — our zero completed free videos turned out to be a screen that never offered the path.
The price × customers arithmetic was available on day one. At $19–49, $1M ARR needs 1,700–4,400 customers, and we had no distribution. It did not require six months.
Comparing across channels with unequal baselines is the weakest design there is. Regression to the mean looks exactly like a treatment effect.
We had the wrong buyer. People who want a channel do not reach for a specialist tool; they reach for ChatGPT. And people already running channels have no reason to use ours.

Why we stopped

Revenue density, compounding asset, lock-in, and the path to the first hundred customers all failed. Only two things were strong: videos really did publish through us into someone else's channel, and better models would lift our quality for free. Even if the remaining experiments had succeeded, the failing columns would have stayed failing. We could have pushed longer, but that buys time, not an answer.

We are now building BuySign — make a webtoon by talking, even if you cannot draw, and collect signs from the people who read it.