The channel ran eight days, put out 26 public videos and stopped. This is what those 508 views actually said — where they came from, which videos held anybody, and the one rule the numbers are strong enough to make canon. Every figure here is pulled from the API, and the two things I could not measure are named at the bottom rather than guessed at.
01 · The run
It decayed to nothing in five days. Sept 13–15 held 79–81% average view percentage. By Sept 17 that fell to 9%, and Sept 20 and 21 drew zero views at all.
That is not the audience leaving. That is the algorithm reading a retention average and switching off distribution — including for the good videos still in rotation. Sept 17 fired three uploads in one afternoon; Sept 18 fired two plus an eleven‑minute explainer that held 15%.
| Day | Views | Watched | Subs | What went out |
|---|---|---|---|---|
| Sep 13 | 89 | 79% | — | — |
| Sep 14 | 139 | 66% | — | 17 videos, one batch |
| Sep 15 | 124 | 81% | — | — |
| Sep 16 | 68 | 50% | +2 | 3 videos |
| Sep 17 | 27 | 9% | — | 3 videos |
| Sep 18 | 14 | 12% | — | 1 long‑form (11 min) |
| Sep 19 | 13 | 44% | — | — |
| Sep 20–21 | 0 | — | — | nothing |
02 · Where the views came from · the surprise
More than half of everything came from people typing a query, not from the Shorts feed. That is unusual at this size, and it is the single fact that should decide how titles get written.
Search brings twice the volume at half the attention. Search visitors watch 41% and last 20 seconds. Feed visitors watch 82% and last 22 seconds on the same videos.
The feed audience likes this content. The search audience is arriving on something that is not quite what they asked for — which is a title problem, not a content problem. And every one of the eight queued shorts was uploaded with an empty tag list, on the channel where search is the main door.
03 · What actually held people
| Title | Views | Watched |
|---|---|---|
| A radio station run entirely by a text file | 42 | 194% |
| Nvidia Bought Groq Without Buying It | 29 | 118% |
| GPT‑6 Shipped. My Workflow Didn’t Change. | 79 | 108% |
| This video was rendered by a text file | 11 | 102% |
| The 5‑line file that stopped my AI writing like everyone else | 6 | 94% |
| New York Banned Hiding AI Faces in Ads | 63 | 48% |
| Tech News Part 10 — Rockwell Builds Smart | 16 | 30% |
| A webhook builds my guest list while I sleep | 12 | 30% |
| Claude skills explained — the four‑line SKILL.md | 6 | 15% |
Proper nouns win the click. Text‑file content wins the watch. “New York Banned Hiding AI Faces” pulled 63 views and lost half of them in thirteen seconds. “A radio station run entirely by a text file” pulled 42 and was watched nearly twice through.
The studio’s own subject is what people rewatch, and it has been treated as the weak lane. The title that does both is already in the library: GPT‑6 Shipped — a named thing for the click, our thing for the watch.
Long‑form has a real verdict now, not a missing one. 15% watched, out at 98 seconds of 669. People arrived and left at a minute and a half. No more eleven‑minute explainers until a short proves the topic pulls.
04 · The rule this produced · canon
A title is not a label on a file. On this channel it is the query it has to match.
— The Line Nobody Read,
(It Applies to You), It Finishes the Job.Part N anywhere. Matches nothing a human types and spends
the first eleven characters saying so. 16, 5 and 2 views.94% of Creators… and
68% of Creators… are the bottom two videos out of 34.05 · What changed · and what is still waiting
The rule lives in the one place every upload already passes through, in
gmcarl‑events, rather than in the pipeline’s head. It
strips a Part N prefix and the Part N is next line
automatically, passes tags through, and warns about the editorial
ones — a title opening on “My”, a bare percentage, a
description with no question in it. It never refuses: a refused upload posts
nothing, and two dark days cost more than one flat headline.
Lifted out of the shipped source, not re‑typed, so the test fails if the worker drifts. It earned its keep immediately: the first version of the “My” rule flagged GPT‑6 Shipped — the best title on the channel — because “My” sits in its first three words. The problem was never the word, it is the word as the subject. Rule narrowed to the opening word only.
Tech News Part 11–18 is gone. Every new title is drawn from
that video’s own description — no invented claims — and each
carries a question written next to its story and 8–10 tags where there
were none. All eight are still private with their publishAt intact,
so they fire Sept 22, 24 and 29 as scheduled. Old snippets saved for a
one‑call revert.
The forward calendar stacked three slots on most days, which is what produced the 9% day. 59 of 62 rows moved and no day now carries more than one: 7 slots left in September, 22 in October, 21 in November. The respread pushed twelve into December, and those were pruned — every one an empty placeholder with no hook, no body and no asset, so it dropped slots rather than ideas. The season now ends when the season ends.
06 · What the numbers cannot tell you
Enabled mid‑review. The YouTube Analytics API was switched on during this
pass, which is why retention exists here and did not an hour earlier. The broad
youtube scope already covered it — no reconnect was needed.
Windows. Retention and traffic sources since 2026‑06‑01; the daily
table since 2026‑09‑01; view totals are lifetime. Analytics counts 470 views
in its window against 508 lifetime, and the gap is the 2021 orphan plus
pre‑window views.
Not measured. Impressions and click‑through are not in this pull, so
“the title won the click” is inferred from views by source, not proven by
CTR.
One comparison is not clean. The radio‑station pair ran on different days.
The weaker framing had the quieter day, so if anything the gap is
understated.
A 200 is not a write. Two of the eight retitles came back 200 OK
with their tag lists silently dropped — title, description and question all
persisted, tags did not. An identical second call stuck. So every write here is
read back and checked rather than trusted, and anything touching this API should do
the same.
n is small. 26 videos and 508 views. The Sept 14 batch is the strongest signal
because seventeen videos shared one day’s conditions and differed only in
their titles.
07 · What happens next
1. Three go out Sept 22 under the new titles, with tags and a question. That is the first clean test of the rule.
2. Pair the lanes. Named actor for the click, the file for the watch. Both halves already exist on this channel and have never been in the same title on purpose.
3. Fifty slots is the runway. One a weekday to Nov 30, and every one of them is an empty title waiting on a story. The constraint is no longer calendar space, it is what goes in the slot.
4. Watch search retention, not views. 41% is the number the title rule is trying to move. Views will bounce around on eight videos; that percentage is the scoreboard.
5. No long‑form until a short earns it. Cut the SKILL.md explainer to 28 seconds with a named hook. If the topic pulls there, it will carry eleven minutes.