Launch Videos

What works in a launch video, and how sure we are

Nine scripting tips from 600 launches, ranked by how useful they are. Every figure is a median with its n. None of it is causal: there is no follower or audience data, so a big company’s reach can explain a big number.

A launch script to test

The tips below, assembled into one runtime. A creative starting point, not a statistically established optimum. For a sub-30-second cut, compress the proof and the ask into one task.

  1. 0–5s

    Hook

    “Watch [specific task] go from [input] to [result].” Show the input and start the action. Or state one verifiable benchmark.

  2. 5–30s

    Name and show

    “This is [product], from [company]. Here’s how it handles [task].” Show real steps, not a feature list.

  3. 30–90s

    Proof

    Show the finished output, one important constraint, and a second use case only if needed. Explain the benchmark’s conditions. Cut this when the proof is already complete.

  4. End

    One ask

    “Try [product] at [URL]” or “Book a demo.” One action, then stop. Track conversions separately from views.

  1. 1

    Open with an actual task when the product can demonstrate one

    Openings that narrate a real interaction (“Create a yellow circle there.”) have the highest median of the common spoken-opening types, more than double a bold-claim opening.

    How sure: Descriptive. It supports testing a task opening, not a claim that a demo in the first five seconds causes views.

    Demo-first launches
    Median views
    1. demo narration792kn=45
    2. stat/number744kn=38
    3. founder intro422kn=52
    4. bold claim324kn=148
    5. problem statement262kn=77

    Library median 407k (n=598)

    stats.json by_dimension.opening_type

    Show the data
    GroupMedian viewsVideos (n)
    demo narration792,00445
    stat/number744,01838
    founder intro421,83452
    bold claim323,885148
    problem statement261,78477
  2. 2

    Use a specific, defensible benchmark instead of a generic superiority claim

    Benchmark/stat hooks have the highest median of any hook style. They appear in 14 top-quartile videos vs. 7 bottom-quartile ones.

    How sure: The repeated-company check weakens this: within companies that launched more than once, benchmark videos sit at 0.91× their own median (n=6). Never invent a number to fit the pattern.

    Benchmark launches
    Median views
    1. benchmark/stat1.2Mn=31
    2. bold claim532kn=50

    Library median 407k (n=598)

    stats.json by_dimension.hook_style

    Show the data
    GroupMedian viewsVideos (n)
    benchmark/stat1,233,02631
    bold claim532,01450
  3. 3

    Follow a claim with evidence the viewer can inspect

    Claim → demo sequences have a median far above a claim with nothing after it, and adding an ending CTA keeps most of that.

    How sure: Samples are small, and some sequences will be mislabelled.

    Bold-claim launches
    Median views
    1. claim → demo1.2Mn=23
    2. claim → demo → CTA792kn=19
    3. claim350kn=158
    4. claim → CTA175kn=44

    Library median 407k (n=598)

    stats.json by_dimension.structure

    Show the data
    GroupMedian viewsVideos (n)
    claim → demo1,230,37623
    claim → demo → CTA792,00419
    claim350,421158
    claim → CTA174,95944
  4. 4

    Choose a compact proof or a developed walkthrough; skip the universal runtime rule

    Sub-30-second videos and 2–3 minute walkthroughs both beat the common 1–2 minute length. There is no single best runtime.

    How sure: The short-video advantage disappears in the small repeated-company subset (0.59×, n=5).

    Median views
    1. Under 30s620kn=41
    2. 30–60s355kn=137
    3. 1–2 min348kn=290
    4. 2–3 min574kn=92
    5. 3 min+344kn=37

    Library median 407k (n=598)

    stats.json by_dimension.length_bucket

    Show the data
    GroupMedian viewsVideos (n)
    Under 30s619,87741
    30–60s354,642137
    1–2 min348,420290
    2–3 min574,31792
    3 min+344,01737
  5. 5

    Identify the company early enough to orient the viewer

    Videos that say the company name in the first 15 seconds have higher medians than those that wait until 15–30 seconds.

    How sure: Only the company name counts, not aliases or product names, so don’t read an optimal second into it.

    Product-name reveals
    Median views
    1. 0–5s546kn=81
    2. 5–15s753kn=43
    3. 15–30s158kn=42
    4. 30–60s300kn=43
    5. after 60s290kn=28

    Library median 407k (n=598)

    stats.json by_dimension.time_to_company_name

    Show the data
    GroupMedian viewsVideos (n)
    0–5s546,21281
    5–15s753,32043
    15–30s158,09142
    30–60s300,46943
    after 60s290,22028
  6. 6

    Write for a clear demonstration, then test narration density

    Both sparse narration (under 80 words a minute) and dense narration (180+) beat the middle. Faster isn’t better, and less copy isn’t better.

    How sure: Pace covers the whole runtime, pauses included.

    Median views
    1. under 80 wpm572kn=84
    2. 80–140 wpm346kn=185
    3. 140–180 wpm327kn=215
    4. 180+ wpm670kn=71

    Library median 407k (n=598)

    stats.json by_dimension.wpm_bucket

    Show the data
    GroupMedian viewsVideos (n)
    under 80 wpm571,51184
    80–140 wpm346,258185
    140–180 wpm326,715215
    180+ wpm670,38071
  7. 7

    Use one concise CTA for the conversion goal; don’t expect it to lift views

    Videos with a detected ending call to action have a lower pooled median than those without one.

    How sure: No conversion data exists here. Keep the ask clear and measure sign-ups or demos separately.

    Median views
    1. no CTA detected428kn=450
    2. CTA detected265kn=105

    Library median 407k (n=598)

    stats.json by_dimension.cta_presence

    Show the data
    GroupMedian viewsVideos (n)
    no CTA detected427,539450
    CTA detected264,878105
  8. 8

    Benchmark against videos in your own category

    Category medians differ fivefold. A model launch is not a target for an enterprise SaaS launch.

    How sure: Audience size and fame differ by category. Don’t change your category to chase a median.

    Median views
    1. hardware/robotics736kn=33
    2. AI model729kn=48
    3. dev tool465kn=76
    4. AI agent377kn=164
    5. fintech212kn=48
    6. enterprise SaaS142kn=61

    Library median 407k (n=598)

    stats.json by_dimension.category

    Show the data
    GroupMedian viewsVideos (n)
    hardware/robotics735,61233
    AI model728,55648
    dev tool464,90076
    AI agent376,795164
    fintech212,04048
    enterprise SaaS142,33061
  9. 9

    Make the first five seconds concrete, and test the wording instead of imposing a formula

    A claim or a number in the first five seconds beats a problem statement. Only eight openings match the strict demo vocabulary that early, and they don’t reproduce the demo advantage.

    How sure: Timings are approximate, and some demo openings are missed.

    Problem-led launches
    Median views
    1. Claim or number441kn=132
    2. Intro or brand459kn=41
    3. Demo cue377kn=8
    4. Problem statement145kn=14

    Library median 407k (n=598)Light bars: fewer than 10 videos

    stats.json by_dimension.first5_content

    Show the data
    GroupMedian viewsVideos (n)
    Claim or number441,103132
    Intro or brand459,05741
    Demo cue376,6508
    Problem statement145,48814

Top quartile vs. bottom quartile

How often each feature appears in each tail149 launches per tail. Bottom median 40k views, top median 2.4M.

Top quartileBottom quartile

  1. Opening: demo narration176
  2. Hook style: benchmark/stat147
  3. Hook style: problem-agitate3456
  4. Length under 30s106
  5. Ending CTA detected2438
Show the data
FeatureTop quartileBottom quartile
Opening: demo narration176
Hook style: benchmark/stat147
Hook style: problem-agitate3456
Length under 30s106
Ending CTA detected2438

Top versus bottom quartiles by views.

Caveats

  • Reach multiple and engagement rate have zero valid observations, so nothing here is normalized for follower count. One video has 138M views.
  • Hook style, opening type and category are AI-assisted labels, not hand-reviewed. Demo, claim and call-to-action cues miss some moments.
  • The corpus is a selected set of launches, with no impressions, spend, retention or conversion data. Tips are hypotheses to test.
  • Within-company checks are small (43 companies with repeat launches, 94 videos) and weaken the benchmark and short-runtime findings.
Weekly, free

Get weekly insights on every new launch video

Every new launch of the week, scored against these patterns.

Hook breakdown included

One short email a week.

Unsubscribe with one click, any time.

You are on the list. The first email arrives next week.

We use your email only to send this digest. See our privacy policy.

Videos belong to the companies that posted them and play through X’s official embed, which loads only after you click. Hook labels and “why it worked” notes are AI-assisted analysis and can be wrong. Curated from public launch posts on X. Data as of Nov 4, 2026. Corrections and removal