How every part is made

A: a hypothetical whole-world event (Yellowstone, Gulf Stream, solar storm). B: a brand the viewer has bought (snacks, fast food, soda, chains, toys, razors), never tech, finance or deals.

The Last Day: whole-world what-ifs 4/7 hits vs 1/9 flops. Bagholder: all 8 tech/finance videos 265-945 views, all 11 consumer-brand videos 136K-1.26M (18, 20, 14)

Copy the two templates exactly. A concrete thing plus a missing piece, never a vague question.

Headline tests: a concrete anchor plus a gap wins the click (n = 22,743 tests). The "deserve" title keyword alone picked Bagholder hits 0.65, better than our judge (05, 15, 19)

Win the click, then keep them: watch time and "was this worth it" surveys rank the video. Upload time and tags don't matter.

YouTube's own statements and interviews (01)

The biggest drop is at the start. A: open inside the event (a news bulletin), "you / everyone" at stake by 60 s, the question by 90 s. B: the anchor number by 30 s, the question by 90 s, the first payoff can wait.

The Last Day hits: cold open 6/7 vs 3/8, stakes 4/7 vs 1/9. Veritasium first chapters run 2:00-4:16 (18, 04)

A visible countdown or numbered chapters (Day 1 to Day 4, Hour 1 to Year 10; rise / bet / fall). Chapters escalate. Never 3 minutes without a new turn.

The one beat that replicates: hits 21/24 vs flops 9/19, p = 0.006 (18, 20)

One big peak about 70% in (copy the source's position). End on a calm reality check or verdict, not an abrupt cut.

Moon peak 72%, Yellowstone 72%. 0 of 35 coded videos end abruptly. Peak-end shapes how people rate the video afterwards (18, 05)

A 134 words a minute, B 154 while speaking. A cut every 4.3 s (A) or 3.1 s (B).

Measured on both sources (16, 17)

A licensed human-actor voice from the ElevenLabs library, never a famous stock voice (Adam). No mispronunciations; vary the delivery. Your pick from the blind test page.

Viewers name a flat, recognisable AI voice as the turn-off; no study compares voices on watch time (08)

A: sound never stops; bed about 12 dB under the voice; it swells about 6 dB in the disaster chapters and changes at chapter starts. B: quieter bed about 17 dB under the voice; a 0.4-0.8 s near-silence at each chapter change, then a new track.

50 ms loudness frames of the Moon and Pepsi videos: Moon 0% silent frames, floor up from -26 to -20 dB in Day 3-4; Pepsi's 4 near-silences land on its 4 chapter changes (351 s, 709 s, 933 s, 55 s). Big channels license 12-70 tracks a video (Fern). Viewers almost never name music (0 of ~330 comments): it is felt, not noticed (02, 04)

A: moving AI clips, no text on screen. B: stock footage, old ads, news headlines with a yellow highlight.

24-frame sheets + cut counts on both sources (17)

Entertainment first. The fact-check only removes what is false. A doubtful detail is swapped for an equally vivid true one or cut; no spoken "reportedly". Sources go in the description.

Gurleen 2026-10-27

Moving clips for every shot (A about $109-172 a video), a licensed actor voice, a real music bed and sound design. Not more script judging: no model could pick hits from scripts.

Kurzgesagt 1,200+ hours and Fern 500-1,000 hours a video go to visuals and sound (04); judge 0.56 and 0.48 (19)