Single Video Deep Analysis (Advanced)

Diagnose why a video underperforms and produce a paste-ready fix kit. Retention and CTR are the primary levers; metadata is secondary. Always ask for the Studio CSV export if the user wants retention diagnosis — never guess Tier 1 numbers.

Data Collection (in order)

1. WebFetch the watch URL → title, description (partial), view/like,

duration, upload date, channel, visible chapters

1. yt-dlp (scripts/fetch_video.py or direct call) → full JSON

metadata, tags, full description, chapters, thumbnails, captions

1. YouTube Data API (if key) → `videos.list?part=snippet,statistics,

contentDetails,topicDetails,status,player,liveStreamingDetails and captions.list + captions.download` (if OAuth)

1. Transcript: prefer manual captions > auto-captions > Whisper

transcription from audio stream

1. Studio CSV (user-provided): retention curve, traffic sources, CTR

by source, impressions, audience tab, real-time first-24h curve

1. Audio loudness: `ffmpeg -i audio.m4a -af loudnorm=print_format=json

-f null -` → integrated LUFS, true peak

1. Thumbnail file: download for analysis via scripts/analyze_thumbnail.py
2. SERP grid: fetch top-10 for primary keyword, save competitor

thumbnails and titles for differentiation scoring

Analysis Dimensions

1. Retention Curve (highest weight when Studio data provided)

Diagnose from the Studio retention curve:

• 0-15s intro: target ≥70% still watching. Below → hook problem.

Diagnose: weak first sentence, no payoff preview, long logo animation, re-introducing yourself ("Hey guys welcome back..."), asking to subscribe before value delivery.

• 15-60s premise: target ≥60%. Below → premise unclear or mismatch

with title/thumbnail.

• Retention cliffs: any drop >10% in <5s is a structural issue. Map

cliffs to transcript timestamps and identify:

• Tangent / digression
• Ad or sponsor break placed badly
• Pacing death (long, slow exposition)
• Promise broken (title said X, video now does Y)
• Visual monotony (static shot >20s without B-roll)
• Mid-video sustain: target curve slope ≥-0.3%/sec. Steeper = boring

middle.

• Ending: the last 30s often rises (re-watchers, end-screen hover).

If it drops hard, end-screen is poorly placed or content feels done before the promise delivered.

• Spike detection: peaks = rewatched moments = high-value moments.

Reuse as chapter titles, thumbnails, Shorts clips.

• APV vs niche median: score video as ×median. Flag anything <0.8×.

If the curve is not provided, state explicitly that retention diagnosis is unavailable and score only metadata, thumbnail, and engagement proxies.

2. Title (target: 60-70 chars, max 100)

• Length (flag >70 truncation risk, critical >100)
• Keyword placement: primary keyword in first 40 chars for Search;

emotional driver in first 40 chars for Browse (which surface matters?)

• Hook type: curiosity, number, contrarian, benefit, authority, fear.

A video should commit to one dominant hook type that matches surface.

• Emotional intensity: rate 1-5. Below 3 = passive titles that lose

Browse impressions.

• Clickability vs. deliverability: can the video content back up the

title? If no, score down (long-term CTR decay).

• Front-load uniqueness: avoid "My thoughts on", "A quick video about"
• Case consistency with channel brand

3. Description

• Above-the-fold (first 150 chars): MUST include primary keyword,

restate title intent, give a click-for-more reason

• Full length: 1,500-4,000 chars ideal. Flag <500 thin, >6,000 bloated
• Entity coverage: list 8-15 Knowledge Graph entities the video

should mention for topic classification. Check transcript + description for coverage. Missing entities = missed semantic relevance.

• Secondary keywords: 3-5 naturally placed in lines 2-10
• Chapters block: present, first = 0:00, ≥3, each ≥10s, descriptive
• Links: grouped and labeled (affiliate, social, related, resource)
• CTA: subscribe + next-video suggestion + lead magnet
• Hashtags: 3 meaningful at the bottom, first is strongest
• FTC disclosure: flag missing if affiliate/sponsored
• Timestamps for Key Moments rich result: correctly formatted

(M:SS - label or MM:SS label), eligible for Google Search carousel

4. Tags

• Count: 5-15 (flag <5 or >20)
• Tag 1 = exact primary keyword
• Mix: 1-3 broad + 5-8 specific long-tail + 2-3 brand/channel tags
• Total chars <500
• Entity overlap: tags should reference the same entities present in

description and transcript (consistency strengthens classification)

• No tag spam (unrelated popular terms)

5. Thumbnail

Delegate to youtube-seo-thumbnail for full treatment. Include a quick score here:

• Resolution 1280x720, <2MB
• Readable at 120px wide (mobile test)
• Face + strong emotion present (niche-appropriate)
• Text ≤4 words, ≥80pt effective
• SERP differentiation (pattern-break vs top-10)
• Consistency with channel grid
• No title-thumbnail redundancy

6. Chapters & Key Moments

• ≥3 chapters
• First is 0:00
• Each ≥10s
• Descriptive, keyword-rich labels (not "Intro / Main / Outro")
• Matches actual structure
• Schema check: propose VideoObject with hasPart[].Clip and

SeekToAction entries so the embed/page unlocks Google "Key Moments" rich result (see optimize skill for code)

7. Captions & Transcript

• Manual captions present (flag auto-only)
• Language tag correct
• Translated captions for top 3 audience countries (from Studio → Audience)
• Transcript contains primary keyword in first 60s and final 60s
• Transcript contains ≥5 Knowledge Graph entities from the topic cluster

8. Audio Loudness

Run FFmpeg loudnorm analysis. Flag:

• Integrated loudness outside -14 to -13 LUFS (quiet = retention drag)
• True peak > -1 dBTP (clipping risk)
• Loudness range >15 LU (inconsistent — causes volume hunting)

9. End-screens & Cards

• End-screen uses all 4 slots (best related, playlist, subscribe, channel)
• Cards at moments of attention (peaks in retention curve if available,

else 60-70% mark)

• Pinned comment: CTA that drives replies (question prompt)
• Pinned comment reply rate (from Studio if available)

10. Engagement Signals

• Like-to-view ratio vs channel median
• Comment velocity (comments / views / day since upload)
• Creator reply rate in first 24h
• Shares, saves-to-playlist visible
• Subscribe-per-1000-views (from Studio)
• Pinned-comment replies

11. Technical & Safety

• Resolution ≥1080p
• Made-for-Kids flag matches content (flag mismatch)
• Altered/synthetic content disclosure if AI-generated
• Copyright claims present? (from Studio)
• Embedding enabled
• Comments enabled with moderation
• Default language + audio language tags set

12. Shorts (if applicable, <60s vertical)

If the video is a Short, apply Shorts-specific rules INSTEAD of the long- form retention analysis:

• 0-1s hook (swipe-away rate)
• Loopability (does the end tie back to the start?)
• Caption overlay for muted viewing
• Audio: trending library audio or original remixable?
• Title ≤40 chars, front-loaded emotion
• `#Shorts` present
• Length sweet spot: 15-30s for loop rate, 45-60s for watch time
• CTA to long-form via pinned comment or end card

Output

Video Score Card

Overall Score: XX/100   (vs niche median: X.Xx)

Retention:         XX/100  ████████░░   [requires Studio CSV]
Title:             XX/100  ██████████
Description:       XX/100  ███████░░░
Tags:              XX/100  █████░░░░░
Thumbnail:         XX/100  ████████░░
Chapters:          XX/100  ██████░░░░
Captions:          XX/100  ███████░░░
Audio Loudness:    XX/100  ████████░░
Engagement:        XX/100  ███████░░░
Technical/Safety:  XX/100  █████████░

Root-Cause Ranking

Rank the 3 largest drags on performance in order, each with estimated impact (e.g., "Hook problem — fixing 0:15 retention from 55% → 70% adds ~20% AVD, estimated +30% impressions").

Issues Found

Critical → High → Medium → Low with one-line fixes and Tier reference.

Paste-Ready Rewrites

Hand off to youtube-seo-optimize for full rewrites OR inline:

• 5 title alternatives (tagged Browse-primary / Search-primary)
• New description body (with chapters, entities, hashtags)
• Tag list
• VideoObject + Clip + SeekToAction JSON-LD schema block
• Pinned comment CTA

Measurement Plan

Error Handling