The claim the show made about itself
The first Empty Folder episode said, on screen, as its honest part: “Nobody has watched any of it yet.”
- Incorrect claim: Nobody has watched any of it yet.
- Correction: Put in front of people 2,115 times. 13 clicked.
Content,Video title,Video publish time,Duration,Views,Watch time (hours),Subscribers,Impressions,Impressions click-through rate (%) 0vgEDOry3NY,An AI Agent Built a Whole YouTube Channel. Here's What Broke. | Empty Folder · Ep. 1,"Aug 4, 2026",636,24,0.5213,0,2115,0.61
There was a dashboard, and it was read every week. It could not see either number. So for a month the honest answer and the true answer were opposites.
This is the month spent measuring the wrong thing, and the three times the same mistake did it.
The same two presenters, and few typed lines
The disclosure, same as always: Pip and Bo are not the people who made this. They are what got made. Every human decision in the story arrives as a typed line, and this month there were very few of them.
A month of measurement
When the episode was made, a month had produced 23 daily snapshots of one channel, not one of them missing; two manual exports from the one screen the automation cannot reach; and two channels, four weeks each, with a number on both sides of the click for the first time.
One correction before any of it counts. An impression is not a person. It is one time a thumbnail was on a screen, and the same person scrolling past twice is two.
| Impressions | Click-through rate | Views | Subscribers | |
|---|---|---|---|---|
| Empty Folder, first episode | 2,142 | 0.65% | 26 | 0 |
| Pipedia, the whole channel | 1,061 | 2.64% | 65 | 1 |
Half the impressions, and 2.5 times the views. One of those got a launch: posts on six networks and around 1,830 impressions of social. The other is a children’s channel that has never been promoted anywhere, on purpose. It is the second row.
The comparison is not a controlled one, and should not be read as one. The personal channel was a year old with 41 subscribers and a back catalogue; Pipedia launched on 29 July from nothing, and new channels are shown to fewer people. The two windows are comparable four weeks, not identical ones.
The dashboard you have is not the one you need
One thing before the detail, because it is the part that transfers. The dashboard you have is not the dashboard you need, and you will not notice, because the one you have answers some question beautifully.
The scheduler’s dashboard answered “how did the post do” every week for a month, fluently, with charts. The question that mattered was “did anyone get the chance to click”, and impressions and click-through rate live on a different screen entirely: YouTube Studio.
A tool that is wrong is easy. A tool that is right about the wrong thing will keep you for a year.
What changed: a checkbox and a naming rule
The process changed in exactly two places.
One. A person opens YouTube Studio, turns on two columns that are off by default — impressions, and impressions click-through rate — and exports a CSV. Those two columns are the entire episode, and being off by default is why it took a month.
Two. Each export goes in a folder named with the date it was taken and the window it covers, because a report for a closed month keeps changing for days afterwards. Two exports of overlapping weeks, taken a week apart, are different data, not a duplicate.
2026-08-16-content-2026-07-19_2026-08-16 2026-08-26-content-2026-07-29_2026-08-26
That is the whole change. No new tool and no new code: a month of not knowing, fixed by a checkbox and a naming convention.
The wall, back from last time
The tool wall from the first episode came back, and this time the diff meant something. One row is new: YouTube Studio, the only source of truth. And one row’s job description was a lie.
- Incorrect claim: Metricool — schedules to YouTube
- Correction: Metricool — schedules. Does not measure
It was never a measurement tool. It was used as one for a month because it was the tool that was open. Nothing on that wall was wrong; the reading of it was.
Lesson 1: a ratio does not fall when its denominator stops arriving
The cost. The only experiment of the month.
The evidence. The first episode’s thumbnail converted at 0.61%, while the same channel’s other videos in the same window ran at 2.00% and 3.72%. So the thumbnail was rebuilt and swapped on the live video on 16 August. Same video, same title, same description: one variable. Textbook.
Nine days later the click-through rate read 0.65%, which looks like a small, honest, boring non-result. It was one click added to 27 impressions.
| | Impressions | Clicks | Views | |---|---|---|---| | v1 (~07-29 → 08-16) | 2,115 | 13 | 24 | | v2 (08-16 → 08-24) | **~27** | **~1** | 2 |
YouTube had stopped showing the video about a week before the thumbnail changed. 22 of its 26 views had landed in one five-day burst, from 4 to 8 August, and the burst was over. The swap was not testing a worse thumbnail against a better one; there was nothing on the other side of it. Telling 0.61% from even 2% needs something on the order of 1,000 impressions on the new version, and it got 27.
And the rate never twitched, because it cannot. Divide by a number that stops arriving and you get the same answer forever.
The rule. Before you read the outcome of any before-and-after, read how much traffic the after got. Exposure first, outcome second, every time.
Lesson 2: accurate, fluent, and pointed at the wrong question
The cost. A month, and it is the reason for the first lesson.
The evidence. For four weeks the scheduler said the first episode had 24 views and no row in its video table at all. That reads as: nobody was ever shown it, so make a better video. The true reading was that 2,115 impressions had produced 13 clicks, so make a better thumbnail.
Opposite diagnoses, opposite fixes, same data. And nothing in the tool was broken. It answered its own question correctly every single week.
The rule. A tool can be accurate, fluent, and pointed at the wrong question.
Lesson 3: blank is not zero
The cost. Nearly a false collapse, which is the only reason it can be described.
The evidence. The daily capture has a reporting lag of exactly three days. A date first appears in the snapshot taken three days later and never changes after that — 23 snapshots at the time, no exceptions. So the last three days of every snapshot are empty. Empty, not zero.
A review was booked for 31 August. It would have read through the 28th and shown three blank days at the end of the window: an apparent collapse immediately after the best week the channel had had.
Same month, same shape, second time: a denominator that was not there, read as a denominator that was zero.
The rule. A gap in a report is a gap in the report.
Lesson 4: two sources are a measurement
The cost. It was only caught because there were finally two sources.
The evidence. At channel level the scheduler and YouTube Studio agreed to within one view across the month. Per video, the scheduler was about a third low: 13 views where Studio said 20, and 3 where it said 5. And the undercount was worst on exactly the two videos that were picking up traffic, which is the cruel version: the error is largest wherever the interest is.
Neither source could have shown that alone. The disagreement is the finding.
The rule. One source is a number. Two sources are a measurement.
Lesson 5: the number you fix becomes the number you watch
The cost. The one lesson nobody wanted to notice.
The evidence. A rate that could not move. A blank read as a nought. A count that was low where it mattered. Three faults in one month, and all three are the bottom of a fraction. The top of each one was watched — clicks, views, rate — and never the exposure underneath.
That is not a coincidence. The diagnosis was about click-through rate, so click-through rate became the thing on the chart, quietly and without anyone deciding it. If impressions per day had been on the chart in the first week, all three faults would have been visible in the first week.
The rule. The number you fix becomes the number you watch. Watch the thing you are dividing by; it is the one that fails silently.
Three more: confident notes, views and dates
The confident note was the wrong one. A tripwire was set: if the traffic-source data was still missing by 17 August, the connection needed reauthorising. It filled in on the 18th, on its own, with nothing reauthorised. The earlier, hedged guess had been right: the account had simply been too small for YouTube to report on. When it arrived it said 52% of traffic came from suggested videos and none from outside YouTube.
Views and clicks are not the same audience. 26 views against about 14 clicks means roughly a dozen views arrived from somewhere that was never an impression at all. A click-through rate and a view count are not two views of one thing; do not put them in the same sentence and call it a funnel.
Every number has a date. The platform keeps revising closed months for days afterwards, so nothing here is the figure. Each one is the figure as of the export it came from.
The control channel found an audience
The part still open when the episode was published was good news breaking something.
On 17 August YouTube started suggesting the children’s channel to people. Impressions went from 342 to 1,061, and the click-through rate did not move.
Content,Video title,Video publish time,Duration,Views,Watch time (hours),Subscribers,Impressions,Impressions click-through rate (%) Total,,,,22,0.3076,1,342,2.63 Total,,,,65,1.1785,1,1061,2.64
That is the one unambiguously good result of the month. Widen the funnel and the packaging usually falls apart; this held at 2.64% across three times the traffic.
Except that channel was the control. It was supposed to sit still and be boring, so there would be something to measure a campaign against later, and it found an audience during its own control period. The baseline a month was spent building was already not a baseline. That was the next question, and it did not have an answer yet.
The dated exports behind every figure
The exports behind every figure here are dated and kept. The project’s write-up now includes the month where it found out how much of what it knew had been guesses. The channel that was never promoted is Pipedia, on YouTube.