The perfect script here means the sequence the most scripts share. Across 82 generative AI startup launch videos from 73 companies, uploaded between 12 June 2024 and 3 August 2026, the first word lands inside a second, the problem by 0:07, the product's name by 0:12 and a trust line near 0:28, in a script that runs a median 1:14. The beat most scripts skip is the ask: 54 of 82 end on a sign-off, an availability line or nothing, Anthropic, Zendesk and Perplexity among them.
313 library videos with a transcript were screened against a rule written before intake; 82 qualified, from 73 companies. Every one is a launch, product or explainer video with a scripted voiceover or presenter, published by a private company founded in 2005 or later whose product is built on generative AI or AI agents, runs 0:20 to 4:00, and was uploaded to YouTube between 12 June 2024 and 3 August 2026; 63 of the 82 were uploaded in 2026. Each script was timed from YouTube's word-level captions.
ADVIDS Research · 82 scripts timed word by word · 73 companies · 14197 caption words · published 12 September 2026 · 19 min read
Top takeaways
The voice starts inside the first second
The first word arrives inside a second in 58 of 82 generative AI startup launch scripts; only 8 wait past three seconds, including Airia, Ataccama, Aviso, etc.
Two in three scripts open on a claim or the product's name, not a question
60 of 82 scripts open on a claim or the product's name; 10 open on a question or a what-if, including Adwisely, Blink, Lusha, etc.
The problem is on the table by 0:07 in most scripts
The problem is stated in 70 of 82 scripts, at a median 0:07; 12 never state one, including ElevenLabs, Shoplazza, AssemblyAI, etc.
The product is named by 0:12, and as often before the problem as after it
The product is named in 80 of 82 scripts at a median 0:12, and in 31 of the 70 scripts that carry both beats the name comes before the problem; scripts that open on it include Amplience, Domino Data Lab, Dragos, etc.
Half the scripts speak a number, and those that do use it by 0:13
A number appears in 41 of 82 scripts, at a median 0:13 when it does; 17 of the 41 speak it inside ten seconds, including Sword Health, Anthropic, AssemblyAI, etc.
The trust line comes late, after the name in 34 of 49 scripts
A trust or control line appears in 49 of 82 scripts at a median 0:28, after the product's name in 34 of 49; 12 carry it inside ten seconds, including Bloomfire, Cube, Drata, etc.
Only 28 of 82 scripts end with an ask
Only 28 of 82 scripts end with an ask; 21 end on a name sign-off, 3 on an availability line and 30 on neither; Anthropic, Zendesk and Perplexity are among the 54.
Questions this report answers
How long is a generative AI startup launch video script?
The median generative AI startup launch video in this set runs 1:14, with the middle half between 0:52 and 1:51, and carries a median 148 words at 127.6 words per minute across 82 videos from 73 companies. 26 of 82 run 60 seconds or less, and 17 run under 0:45, including Bloomfire, Zapier, Airia, etc.
When should an AI launch video name the product?
In 82 generative AI startup scripts the product is named at a median 0:12; 36 of 82 name it inside ten seconds and 11 wait past 0:30. Scripts that speak the name inside the first second include Amplience, Domino Data Lab, Dragos, etc.; Ataccama waits until 1:23 of a 1:38 script.
Do AI product launch scripts start with the problem or the product?
Neither dominates: of the 70 scripts that state a problem and name the product, 39 put the problem first and 31 the name first. Problem-first scripts include Lyzr, Sword Health, Ada, etc.; name-first scripts include Bloomfire, Sword Health, Amplience, etc.
Do AI launch videos open with a question?
Rarely: 5 of 82 generative AI startup scripts open on a question and 5 on a what-if or imagine line, while 37 open on a declarative claim and 23 on the product's name. 53 of 82 scripts ask no question at all.
Do generative AI launch scripts end with a call to action?
Most do not: 28 of 82 end with a direct ask such as book a demo, start building or a URL, 21 on the product's name as a sign-off, 3 on an availability line and 30 on none of these. Scripts that end with an ask include Shoplazza, Ada, Adwisely, etc.
How were these scripts measured?
ADVIDS Research screened 313 library videos against a rule written before intake; 82 qualified, from 73 private companies founded in 2005 or later whose product is built on generative AI. Each script was timed from YouTube's word-level captions, 14197 words in all, and every beat is a published rule over those words.
How to read this report
- The set is 82 launch scripts from 73 generative AI startups. A script is one video's spoken words; a beat is one coded moment in it, such as the problem sentence or the ask.
- "N of M" counts scripts: N carry the value out of a base of M. The base is all 82 scripts unless the sentence names a smaller group, such as the scripts that carry both beats in an order claim.
- A timecode is the moment a beat's first word is spoken. On the 60-second ruler a position is the median share of runtime scaled to 60 seconds, with the actual median timecode beside it.
The 60-second map
Positions are median shares of runtime scaled to 60 seconds; the median script runs 1:14. The ask is drawn at the median position of the 28 scripts that make one.
Each panel is the same 82 scripts, shortest at the top, longest at the bottom; the x-axis runs from the first frame to the last. Problem and name cluster at the left edge, the trust line spreads through the middle, and the ask sits at the right edge in the 28 scripts that make one.
Beat 01 · First word · 0:00 on a 60-second script
1. The voice starts inside the first second
The first word of a generative AI startup launch script lands inside one second in 58 of 82 scripts and inside two seconds in 70; 8 wait past three seconds. The middle half of scripts start between 0.0 and 1.2 seconds. At the other end the voice stops a median 3.8 seconds before the video does, and 20 scripts leave more than five seconds of picture after the last word. The script does not wait for a logo. The logo waits for the script.
The 8 scripts that wait past three seconds include Airia, Ataccama, Aviso, etc., and the 20 that leave more than five seconds of picture after the last word include Airia, Amplience, Aviso, etc. 36 of 82 scripts speak the product or company name again inside the final tenth of runtime, which is where the silent tail sits.
Perplexity holds the picture for six seconds. Its Perplexity Computer explainer opens on interface motion and speaks its first word at 0:06, the latest in the set, then names the product in that first sentence.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
Beat 02 · Opening line · 0:00 on a 60-second script
2. Two in three scripts open on a claim or the product's name, not a question
60 of 82 generative AI startup launch scripts open on a declarative claim (37) or on the product or company name in the first ten words (23). 9 open on a problem sentence, 5 on a what-if or imagine line, 5 on a question and 3 on a number. 53 of 82 scripts ask no question anywhere. The opening line asserts; it does not ask.
Claim openers include Lyzr, Sword Health, Alloy, etc.; JobAdder opens on "Recruitment is moving faster than ever" and Drata on "Businesses run on trust". Name-first openers include Bloomfire, Ada, Amplience, etc., problem openers include Airia, Akto, Ataccama, etc., question and what-if openers include Adwisely, Blink, Lusha, etc., and the 3 statistic openers are Keka, Nooks and Plotly.
Swimlane opens on a question and a joke. "Does your SOC feel out of tune?" leads into an orchestra metaphor and a band of agents, and the script still names the product by 0:29 and ends with an ask at 0:44.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
The question opener that copywriting advice favours appears in one script in eight.
Beat 03 · Problem stated · 0:07 on a 60-second script
3. The problem is on the table by 0:07 in most scripts
70 of 82 generative AI startup launch scripts state a problem, at a median 0:07; the middle half land between 0:03 and 0:26. 40 state it inside ten seconds and 12 wait past 0:30. 12 scripts never state one, and the group leans to developer tools and model launches, including ElevenLabs, Shoplazza, AssemblyAI, etc. The problem sentence is short, early, and optional for the companies selling to builders.
Scripts that state the problem inside ten seconds include Bloomfire, Sword Health, Ada, etc.; those that wait past 0:30 include Sword Health, Cequence Security, Contentstack, etc.
AssemblyAI states no problem at all. Its Universal-3 Pro launch is a specification read: a new class of speech language model, what it captures, what it costs, and an ask to start building, in 0:37. When the buyer already knows the problem, the script spends its seconds on the spec.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
The problem sentence is front-loaded: when it comes, it comes early, and one script in seven never states one.
Beat 04 · Product named · 0:08 on a 60-second script
4. The product is named by 0:12, and as often before the problem as after it
80 of 82 generative AI startup launch scripts speak the product or company name, at a median 0:12; 36 name it inside ten seconds and 11 wait past 0:30. Of the 70 scripts that both state a problem and name the product, 39 put the problem first and 31 put the name first, a median 5.3 seconds apart. 38 of 82 use an introducing, meet or announcing phrase. The problem-then-solution order is a habit, not a rule: the set splits down the middle.
Name-first scripts include Bloomfire, Sword Health, Amplience, etc.; problem-first scripts include Lyzr, Sword Health, Ada, etc.; those that name the product after 0:30 include Adwisely, Akto, Ataccama, etc. Two scripts never speak a name: Replit and Shoplazza; Replit's 0:30 Agent 4 spot is a two-person dialogue and Shoplazza's store-builder spot says only "our free AI store builder".
Ataccama withholds the name until 1:23 of a 1:38 script. The script runs the old world, the data team's problems and the platform's behaviour before "Introducing Ataccama ONE", then asks for a demo. A reveal structure that names late is possible; it is the exception, eleven scripts in 82.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
Naming is front-loaded like the problem, and the two beats sit a median 5 seconds apart in either order.
Beat 05 · Number or proof · 0:09 on a 60-second script
5. Half the scripts speak a number, and those that do use it by 0:13
41 of 82 generative AI startup launch scripts speak a number or a proof phrase, and 41 do not. Where a number appears it arrives at a median 0:13; 17 of the 41 speak it inside ten seconds. In 24 of the 41 scripts with both beats the name comes before the number. The number is a credential said early, or it is not said.
The 17 numbers spoken inside ten seconds include Nooks on thousands of sales teams and billions in pipeline, Plotly on a billion knowledge professionals, Maze on 3,000 companies and Superblocks on 1,000 apps. Scripts with no number include Bloomfire, Zapier, Airia, etc.
Nooks opens on the number. "Thousands of the world's best sales teams are generating billions in pipeline with Nooks" is the first sentence, name and proof together at 0:00, before any problem.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
A number is either an opening credential or absent; it is rarely saved for the middle.
Beat 06 · Trust or control line · 0:23 on a 60-second script
6. The trust line comes late, after the name in 34 of 49 scripts
49 of 82 generative AI startup launch scripts carry a security, privacy, compliance, governance or control line, at a median 0:28; the middle half sit between 0:12 and 0:43. It follows the problem in 33 of the 43 scripts with both, and follows the product's name in 34 of 49. 33 scripts, among them the voice and developer products, carry none. Trust is the third act of a startup script: problem, name, then reassurance.
Trust lines inside ten seconds include Bloomfire, Cube, Drata, etc.; those after 0:30 include Adwisely, Akto, Amplience, etc.; scripts with none include Sword Health, Shoplazza, Ada, etc.
Drata makes trust the opening line. "Businesses run on trust" is its first sentence at 0:02, the product follows at 0:13, and the script closes on "Drata. Win with trust" with no ask. For a compliance product, the trust line is the problem statement.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
Unlike the problem and the name, the trust line is a middle-of-script beat.
Beat 07 · Close · 0:55 on a 60-second script
7. Only 28 of 82 scripts end with an ask
28 of 82 generative AI startup launch scripts end with a direct ask: book a demo, start building, try it free, a URL. 21 end on the product's name as a sign-off, 3 on an availability line, and 30 on none of these. Where an ask exists it lands at a median 91.2 per cent of runtime, 0:55 on a 60-second script and 1:05 in the actual set. The script most teams ship is a claim, a name and a fade.
Scripts that end with an ask include Shoplazza, Ada, Adwisely, etc.; name sign-offs include Akto, Alloy, Anthropic, etc. The 3 availability lines are ElevenLabs, Kentik and Snaptrude, and scripts that end on none of these include Bloomfire, Sword Health, Zapier, etc.
Superblocks asks for a comment, not a demo. Its 3.0 launch ends "Comment a SaaS tool you hate and why, and we'll build an enterprise-grade replacement" at 2:16, an ask that turns the viewer into the brief.
Source: ADVIDS Research coded set, 82 scripts from 73 companies, uploaded 12 June 2024 to 3 August 2026
The most common ending is a line that asks nothing and names nothing.
Benchmarks
Descriptive figures for placing your own script against the set.
Script length and pace
| Measure | Median | Middle half | What it means |
|---|---|---|---|
| Runtime, median | 1:14 | 0:52 to 1:51 | Half the videos run shorter than this; the range is the middle half |
| Words spoken, median | 148 | 94 to 215 | Caption words per script, and the middle half |
| Words per minute, median | 127.6 | 112.9 to 145.2 | Spoken pace over the whole runtime |
| Sentences, median | 14 | 9 to 19 | Sentences per script, split on terminal punctuation |
| Words per sentence, median | 10.9 | 8.9 to 13.2 | Mean sentence length per script |
| First action verb, median | 0:17 | 0:08 to 0:29 (n=69) | First connects, reads, generates, automates or similar; 69 of 82 scripts carry one |
| 'You' per 100 words, median | 3.6 | 1.8 to 5.7 | Second person addressed to the viewer; 65 of 82 scripts use 'you' more than 'we' |
| 'We' per 100 words, median | 0.6 | 0.0 to 1.2 | First person plural; the company speaking about itself |
| Tail after the last word, median | 3.8 | 2.4 to 5.0 s | Seconds of picture after the final spoken word |
Runtime of the 82 videos
| Runtime | Scripts | Base |
|---|---|---|
| Under 0:45 | 17 | of 82 |
| 0:45 to 1:15 | 28 | of 82 |
| 1:15 to 2:00 | 21 | of 82 |
| Over 2:00 | 16 | of 82 |
What most scripts skip
| Beat | Scripts | Base | Median timecode | Companies include |
|---|---|---|---|---|
| Contrast with the old way (no more, without, instead of, unlike) | 33 | of 82 | 0:26 | Ada, Adwisely, Alloy, etc. |
| Introducing, meet or announcing phrase | 38 | of 82 | 0:13 | Sword Health, ElevenLabs, Ada, etc. |
| Audience named (for X teams, if you're a) | 17 | of 82 | 0:14 | Alloy, Ataccama, Cequence Security, etc. |
53 of 82 scripts ask no question anywhere, and the median script asks 0. 9 of 82 never say "you"; 65 use "you" more often than "we". Beats under the half-the-set floor are listed here rather than mapped.
What to do with this
- Beat 07 If 54 of 82 generative AI launch scripts end without an ask, read your last spoken line aloud and decide what a viewer can do with it.
- Beat 01 If 70 of 82 scripts have spoken by 0:02, move the logo bumper to the end and start the voice on frame one.
- Beat 02 If 60 of 82 open on a claim or the name, write the first sentence as an assertion and cut the rhetorical question.
- Beat 04 If the set splits 39 to 31 on problem-first against name-first, choose the order for your audience, not from habit, and name the product by 0:12 either way.
- Beat 05 If 41 of 82 scripts speak a number inside the first quarter, decide which figure the company will say aloud before the script draft starts.
- Beat 06 If the trust line lands at 0:28 in 49 of 82 scripts, and your buyer is an enterprise security reviewer, test it as the opening sentence instead.
Methodology
| Vertical (visual record only) | Scripts |
|---|---|
| devtools | 4 |
| ai | 2 |
| fintech | 2 |
| knowledge_management | 2 |
| sales | 2 |
| security | 2 |
| adtech | 1 |
| compliance | 1 |
| Other | 10 |
| Not recorded | 56 |
| Coding scheme | Values | Definition |
|---|---|---|
| First word | seconds | Timecode of the first caption word. Tail silence is the runtime left after the last word. |
| Opening type | claim · product-first · problem · question · what-if · statistic | The first sentence, classified in this order: the product or company named in its first ten words; a what-if or imagine; a question; a number with a magnitude; a problem cue; otherwise a claim. |
| Problem stated | seconds | First sentence carrying a problem cue: still, manual, slow, chaos, buried, waste, hard, struggle, can't, don't, without, broken, missing, drowning, overwhelmed, tired, never, fail, stuck, friction, silo, scattered, mess, pain, bottleneck, risk, delay, complex, fragmented, outdated, legacy. |
| Product named | seconds | First spoken form of the company or product name, checked per script against the captions (Sana is captioned Sauna, Zendesk as Zenesk). |
| Number or proof | seconds | First figure with a magnitude, percentage, unit or count, or a trusted-by, used-by, thousands-of phrase. |
| Trust or control line | seconds | First of: secure, security, privacy, private, compliance, governance, guardrails, control, your data, SOC 2, GDPR, HIPAA, permissions, audit, trust, encryption, access control. |
| Close | ask · availability · sign-off · none | In the final fifth of the words: a direct ask (get started, try, book a demo, learn more, visit, join, download, contact, a URL, an imperative ending in today); else an availability line (available now, out now, now live); else the name spoken in the last tenth of runtime; else none. |
| Skipped beats | contrast · audience · introducing phrase | Contrast: no more, without, instead of, not just, stop, forget, unlike. Audience: for X teams, if you're a, a named role. Introducing phrase: introducing, meet, say hello to, announcing. |
| Script scalars | words · words per minute · sentences · you and we per 100 words · first action verb | Counts over the caption words. The first action verb is the first of connects, integrates, reads, learns, generates, automates, analyses, detects, monitors, routes, writes, drafts. |
Cite this report
Jai Ghosh, Advids. The Generative AI Video Teardown: We Audited 73 Startups to Map the Perfect 60-Second Script. 12 September 2026. https://advids.co/blog/generative-ai-startup-video-script-map
About Advids
Advids produces video for AI, SaaS and DeepTech companies. ADVIDS Research publishes coded studies of what those companies ship, with the dataset attached so every count can be checked.
Image credits and licensing
Charts and images credited ADVIDS Research, and the coded dataset behind them, are © 2026 Advids and licensed CC BY 4.0: reuse them with attribution to Advids and a link to this page.