Video Breakdown AI – Hook Analysis, Shot List & Script

By TubeText Labs – see all our tools at the end of this page.

Find out why a video works in one API call. Paste TikTok, YouTube/Shorts or other public video links and get a structured breakdown: the hook, a shot-by-shot storyboard, measured cut pacing, the products on screen, whether it's organic or an ad/affiliate post, the sound used and the full script.

Built for UGC and performance marketers, ad buyers, agencies and creators who build swipe files and briefs from what's already winning.

What you get for each video

• 🎣 Hook – what happens in the first ~3 seconds (sampled densely), hook type and why it works
• 🎬 Shot list – start/end, framing, scene, action, expression, on-screen text, spoken line
• ✂️ Measured pacing – real cut detection that finds ~9 in 10 hard cuts in short videos (TikTok/Reels style), including jump cuts: cut count, average shot length, cuts per minute
• 🛍️ Products shown on screen – name, brand when visible, first appearance time, how it's shown (incidental objects and ingredients are left out)
• 💰 Commercial intent – organic / affiliate & link in bio / TikTok Shop / brand ad / self-promotion, with the evidence quoted (it stays "organic" unless there's explicit evidence)
• ⭐ Key moments – the reveal, the "money shot", the objection handler, with timestamps
• 🧭 Why it worked – a short video-level summary of what makes it work
• 🔈 Audio & hook flags – hasSpeech, hasVoiceover, hasMusic, and how the hook is delivered (hookDeliveredAs: text overlay / voice / visual / mixed)
• 🎵 Sound – title and author from the platform's sound info, plus how the audio is used (voiceover, dialogue, music bed, trending sound), judged from the audio track
• 📈 Engagement rate plus views, likes, comments, shares and saves
• 🚀 Outlier score – is this a breakout for its creator? TikTok: likes vs the creator's average likes per video. YouTube: views vs the median of the channel's ~30 latest uploads. isOutlier is true at 3× or more. Plus creator stats for TikTok (followers, total likes, video count).
• 🧱 Structure, CTA, format, tone, audience and the full transcript (captions or Whisper)
• 🌍 Output in English by default or 18 other languages; quotes can stay original or be translated

Use cases

• Build swipe files of winning UGC ads and viral hooks
• Turn competitors' best videos into briefs and scripts for your creators
• Analyze hundreds of videos in a niche to see which formats and hooks dominate
• Feed structured breakdowns into ChatGPT/Claude to generate new script variations

Find videos to analyze (no links needed)

You don't need a list of links. Point the Actor at a source and it picks the videos, then breaks each one down:

• YouTube keyword search – searchQueries (e.g. ["skincare routine"]) with searchPublishedWithin (hour / today / week / month / year / any; default week) and searchSortBy (views / relevance / date; default views, re-sorted by actual view count).
• Today's trending by country – trendingCountries (e.g. ["us","jp","gb","de","br","in","kr","fr"]), from daily YouTube trending lists (via kworb.net).
• A channel's most viewed videos – a channel link plus channelSort: "popular" (or "newest", the default). Works on /videos, /shorts and /streams.
• Top videos from another Actor's results – datasetId: pick the dataset of any TikTok, Instagram or YouTube scraper run. We read the video links and, by default, take the most viewed (datasetSortBy: "views"). This is the way to analyze top TikToks from a hashtag, trending or profile scraper.

Each source is capped by maxVideosPerSource (default 10). Finding videos is free; each analyzed video is billed as usual.

Example input:

{
  "searchQueries": ["skincare routine"],
  "searchPublishedWithin": "week",
  "maxVideosPerSource": 10
}

Video to script

Get the full spoken script with timestamps plus the shot-by-shot structure, so you can rewrite a winning video into a new script or brief.

Real example (TikTok UGC ad, 39.5M plays, 76 s) – shortened v1.1 output

{
  "format": "product demo",
  "hook": {
    "type": "visual surprise",
    "text": "Purple strip applied to a yellow banana – 'yellow + purple = white'",
    "why_it_works": "The absurd use of a banana as a proxy for teeth creates immediate curiosity and a visual pattern interrupt."
  },
  "commercialIntent": {
    "type": "affiliate/link in bio",
    "evidence": "The creator explicitly says 'i'mma leave a link to them' and uses the hashtag #tiktokmademebuyit."
  },
  "products": [
    { "name": "purple whitening strips", "brand": "", "first_seen": 0, "how_shown": "applied to banana and teeth, packaging held to camera" }
  ],
  "keyMoments": [
    { "time": 47, "what": "Bottom teeth visibly whiter than the top teeth", "why_it_matters": "The 'money shot' that provides the visual proof needed to convert the viewer." },
    { "time": 59, "what": "Teeth shown again after brushing", "why_it_matters": "Addresses a common doubt about how long the result lasts." }
  ],
  "pacing": { "cutCount": 5, "avgShotSeconds": 12.7, "cutsPerMinute": 3.9 },
  "engagementRate": 0.0169
}

More results from the same test (10 viral TikToks, 5.6M–39.5M plays):

Video type	What the breakdown found
GRWM makeup routine (14.9M)	21 shots for 20 measured cuts, 26.7 cuts/min; trending song identified from sound info; "result first" hook
Kids' savings tutorial (9.1M)	Affiliate intent from "Link in my bio"; 3 products with first-seen times; 20 cuts
Bizarre beauty product review (8M)	3 products identified; "I lost my eyebrow!" fail moment at 0:09
Gameplay clip (10.4M)	Format "gameplay clip", game name and meme sound from metadata

Accuracy, honestly

Latest independent check (Oct 2026, 22 videos across TikTok, Instagram Reels and YouTube): all 22 were analyzed. Measured cuts matched an ffmpeg scene check exactly on 8 of 9 TikToks. Products and commercial intent were right on all 5 shop videos. Commercial intent was wrong on 2 of 12 non-shop videos; it has since been tightened to require explicit evidence. On-screen text was captured on 17 of 22. Cut detection was then calibrated on 30 TikToks with 515 hand-labelled cuts: it finds 91% of real cuts (including jump cuts), and 96% of the cuts it reports are real. Flashes and strobes are ignored. Known limits:

• Details inside one long continuous shot can be merged into a single shot description.
• A short visual moment can be missed.
• Song titles come only from the platform's sound info (the AI does not identify songs by ear). The audio track is used to judge voice, music and sound effects. Speech comes from captions or Whisper. When the audio is music only, song lyrics picked up by captions or Whisper go into a separate lyrics field instead of the transcript (hasSpeech: false).
• Instagram Reels (beta) are slower, up to about 5 minutes per reel.

Input example

{
  "urls": [
    "https://www.tiktok.com/@creator/video/7682492258666515726",
    "https://www.youtube.com/watch?v=EMvk7OC4OeY",
    "https://www.tiktok.com/t/ZTyXc9cpY/"
  ],
  "outputLanguage": "English"
}

Languages: the analysis is in English by default. Pick another outputLanguage (Spanish, Portuguese, French, German, Japanese, Korean, Chinese and more, or Same as video). Quotes (spoken words, on-screen text) stay in their original language unless you turn on translateQuotes, which is handy for studying foreign-language ads.

Output example (shortened)

{
  "status": "ok",
  "platform": "tiktok",
  "url": "https://www.tiktok.com/@creator/video/7682492258666515726",
  "author": "creator",
  "durationSeconds": 41,
  "summary": "A father humorously reviews his newborn as if the baby were a consumer product…",
  "hook": { "text": "‘2026 June Edition – product review’", "type": "curiosity gap", "why_it_works": "Familiar review format applied to a baby creates instant surprise." },
  "format": "Talking head comedy skit",
  "structure": [{ "beat": "hook", "start": 0, "end": 3, "description": "…" }],
  "shots": [
    { "start": 17, "end": 19, "shot": "medium", "scene": "Living room", "action": "Looks for a ‘volume button’ on the baby; baby starts crying", "expression": "Confused", "on_screen_text": "to figure out", "spoken": "We're still trying to figure out where the volume…" }
  ],
  "cta": "",
  "editingNotes": "Bold captions highlight the ‘technical’ product terms…",
  "transcript": "First time user here and I've had the 2026 June edition…"
}

Pricing

Pay per event:

• $0.03 per analyzed video (up to 3 minutes; cheaper on higher Apify plans)
• $0.01 per extra started minute for longer videos (a 10-minute YouTube video = $0.10), capped at 30 extra minutes, so the analysis costs at most $0.33 per video however long it is (it samples at most 30 frames); optional residential traffic is billed separately
• Free: invalid links, and deleted, private, unavailable, live or too-long videos. Also free if our own analysis fails.
• If YouTube is blocking broadly during a run, the Actor stops retrying and marks the remaining videos youtube_blocking_now. They are not charged; please try again later.

From Oct 22, 2026 (06:56 UTC):

• $0.005 per video we tried but couldn't fetch (for example YouTube bot checks or a proxy timeout). This covers our download cost. Invalid, unavailable, private, too-long and live videos stay free.
• Optional residential proxy for YouTube (proxyMode: "residential"): $0.015 per started MB of residential traffic, charged only if you choose this mode. It's usually 3–10 MB per YouTube video, and it gets through YouTube's bot checks far more often than the default datacenter route. Until Oct 22 this mode falls back to datacenter.

Similar video-analysis Actors on the Store charge $0.04–$0.10 per video (checked Oct 2026).

Supported links

• YouTube: videos and Shorts (youtube.com/watch?v=…, youtu.be/…, /shorts/…), plus channels and playlists (youtube.com/@name, @name/shorts, @name/streams, playlist?list=…). Channels and playlists expand to the newest videos, up to maxVideosPerSource (default 10, max 200). Each video is billed like a single video, and videos longer than maxVideoMinutes are skipped and not charged.
• TikTok: videos, photo slideshows, tiktok.com/t/… and vm.tiktok.com/… share links and embed links; different links to the same video count once
• Instagram: public Reels (instagram.com/reel/…), no login needed (beta)
• Other sites: public video pages supported by yt-dlp, tested with Reddit, Bilibili (Chinese titles kept, analysis in your output language), Facebook public videos and Dailymotion, plus direct video file links (.mp4, .mov, .webm)
• Not supported: TikTok and Instagram profile links directly (paste video links, or chain a profile scraper's dataset via datasetId), live streams in progress, content behind a login (e.g. private Instagram accounts, Vimeo), private or members-only videos, posts without a video (no_video_in_this_post) and DRM-protected videos (drm_protected). These come back as skipped and are not charged

How it works

For each video the Actor detects every hard cut, then samples frames densely in the first 3 seconds (the hook), right after each cut, and evenly across the rest (up to 30 frames). These frames go to a vision-language model together with the timed transcript and the video's facts (caption, hashtags, sound name, stats, measured cuts), and the model returns a structured breakdown. For videos up to ~50 minutes the audio track is sent too, so the model can judge voiceover, music and sound effects; song titles still come only from the platform's sound info, and speech from captions or Whisper.

Long YouTube videos (over ~2.5 minutes). If the video has captions, the Actor uses a fast mode: frames are sampled evenly across the video (no cut detection, no audio track) and speech comes from the captions. If there are no captions, it downloads the full file to get audio, cut detection and a Whisper transcript; if that download is too slow, it falls back to frames only.

Ready-made examples

• TikTok Hook Analysis
• UGC Ad Analysis
• YouTube Trending Videos Analysis
• Viral Video Analysis by Keyword
• YouTube Trending Videos Analysis – UK
• YouTube Trending Videos Analysis – India
• YouTube Trending Videos Analysis – Japan
• YouTube Trending Videos Analysis – Brazil
• TikTok Shop Video Analysis

TubeText Labs tools

TubeText Labs builds 9 pay-per-result Apify Actors for video, social and e-commerce data – no login, JSON/CSV/Excel output, and ready for ChatGPT and Claude (Apify MCP), n8n, Make and Zapier. All tools: apify.com/tubetext

Transcripts

• YouTube Transcript Scraper – videos, Shorts, playlists and whole channels to text; Whisper AI when there are no captions
• TikTok Transcript Scraper – TikTok captions plus Whisper AI, SRT and timestamps
• Instagram Reels Transcript – reel links or @usernames to text with Whisper AI

Video & ad analysis

• Video Breakdown AI (this tool) – hook analysis, shot list, cut pacing, products shown and script of any TikTok, Reel or YouTube video
• Video Ad Analyzer – competitors' YouTube video ads by brand or domain (Google Ads Transparency), torn down by AI

Channel analytics

• YouTube Channel Analytics – subscribers, recent video performance, Social Blade grade, ranks and estimated earnings

E-commerce & travel

• TikTok Shop Product Scraper – US TikTok Shop prices, items sold, stock per SKU and shop stats
• AliExpress Search Scraper – AliExpress search results by keyword: price, items sold, rating, shipping, badges
• Google Hotels Scraper – hotel prices for any destination and dates, per booking site and as a 90-day price calendar

FAQ

Can I paste a channel, profile or playlist? YouTube channels (including the /shorts and /streams tabs) and playlists: yes. You get the newest videos by default, or the most viewed with channelSort: "popular", up to maxVideosPerSource (default 10, max 200), and each video is billed like a single video. For TikTok or Instagram profiles, run any profile scraper and pass its dataset in datasetId.

Should I use the residential proxy mode? Only for YouTube, and only if the default mode returns many failed downloads for your videos. TikTok, Instagram and most other sites work fine with the default. From Oct 22, 2026, residential mode costs $0.015 per started MB (usually 3–10 MB per YouTube video) on top of the per-video price.

Why is outlier empty? It needs a creator baseline: it's null for creators with fewer than 5 videos and for sites other than TikTok and YouTube.

Does it work without speech? Yes. Visuals and on-screen text are analyzed even when nobody talks.

Is this legal? The Actor only reads publicly available video pages, without logging in. You are responsible for how you use the data (copyright, privacy and TikTok's terms).

How do I get a clean spreadsheet? In the run's Storage tab choose Export → View: Overview for a short set of readable columns in Excel, CSV or Google Sheets. Use JSON to get every field (for example the full shot list).

Can I cap what a run costs? Yes – set Maximum cost per run in the run options. The run stops cleanly when the next item would go over it: you are never charged more, and you get a clear "stopped at your limit" status message.

What happens if a big run hits the time limit? It stops cleanly shortly before the limit: everything finished so far is delivered, the run ends as succeeded, and the status message says how many items weren't started. Raise the run timeout, or run again with the inputs that were not started.