← BACK TO APP

HOW IT WORKS

YouTube Intelligence is a full-stack AI application that searches YouTube for any topic, pulls transcripts and channel stats, runs advanced LLM synthesis in the cloud via Gemini, and renders the results in a brutalist web dashboard. The frontend is hosted on Vercel, powered by a serverless Python backend running on Modal.

YouTube is where purchase decisions are made — especially for expensive, considered purchases like cars, appliances, and tech. It is one of the few platforms that exposes reach mechanics transparently: view counts show how far an opinion has travelled, like/dislike ratios show whether it landed, and subscriber counts let you weight enthusiast voices against casual ones. This tool turns that public opinion layer into structured commercial intelligence.

70%
of car buyers say YouTube video content influences their purchase decision. Watch time of "test drive" videos has grown more than 65% in two years, and 64% of shoppers say new video formats like 360° would convince them to buy without a physical test drive.
1.7×
more likely to positively influence brand consideration than TikTok, Instagram, and Facebook — per a March 2025 study of 10,000 US shoppers commissioned by Google from Boston Consulting Group. YouTube content is also rated 1.7× more relevant and 2× more trustworthy than other social platforms by shoppers.
50%
of shoppers where digital video played a role in their path to purchase said it made them aware of products or brands. 45% said it helped them choose which product to buy. 34% said it directly prompted a purchase. Video shapes the full funnel, not just awareness.
Gap
Despite YouTube's outsized influence, it receives a disproportionately low share of ad spend relative to its actual impact on consumer decisions. Brands are systematically underinvesting in the platform where opinions are being formed — and under-monitoring what is being said there.

You enter a search query like "Best fitness watches in Australia". The frontend sends a request to your serverless Modal backend. Modal finds the top YouTube videos for that query, fetches their full transcripts, pulls video and channel stats from the YouTube Data API, grabs the top comments, and runs 5 structured LLM prompts per video (via Google Gemini) to extract sentiment, themes, comment breakdown, creator vs audience agreement, and sponsorship detection. It then runs a cross-video meta analysis to identify temporal trends and unique points of view across all videos.

Optionally, enter a client name on the results dashboard and the app calls your Modal backend again to generate a client-specific intelligence report with prioritised, effort-tagged recommendations.

LayerToolWhat for
FrontendVercel (HTML/CSS/JS)User interface, state management, live dashboard rendering
Backend APIModal (FastAPI/Python)Serverless execution environment for heavy scraping and processing
YouTube Search & TranscriptsSerpAPIFind video IDs and fetch full spoken transcripts per video
Video & Channel StatsYouTube Data API v3Views, likes, subs, dates, duration, top comments
LLM — Intelligence PipelineGoogle GeminiSentiment, themes, and comment analysis per video
LLM — Client ReportGoogle GeminiClient intelligence generation from compiled dashboard context
This decoupled architecture ensures that all heavy processing (which can take minutes) runs securely on Modal's cloud infrastructure. This completely bypasses standard frontend HTTP timeouts and keeps all API keys strictly hidden from the public.

To run the backend, you need an account on Modal.com and three API keys stored securely as Modal Secrets:

1. SerpAPI key

Go to serpapi.com → sign up → Dashboard → API Key

2. YouTube Data API v3 key (free)

Go to console.cloud.google.com Create project → Enable APIs → search "YouTube Data API v3" → Enable Credentials → Create Credentials → API Key

3. Gemini API Key

Go to Google AI Studio (aistudio.google.com) → Get API Key
In your Modal dashboard (modal.com/secrets), create a new custom secret named exactly Youtube-App. Add three environment variables to it: SERPAPI_KEY, YOUTUBE_API_KEY, and GEMINI_API_KEY. Your Python backend will automatically load these at runtime.
SectionWhat it showsHow it's calculated
Channel DistributionWhich channels drive the narrative by share of total viewsEach channel's views ÷ total views. Top 5 shown, rest grouped as Others. Like ratio percentile tags added per channel.
Dominant ThemesWhat topics come up most across all videosLLM cross-video meta analysis. Percentage estimated from frequency across transcripts.
Temporal TrendHow the conversation shifted from oldest to newest videoLLM analyses videos sorted by publish date. No video numbers — natural prose only.
Unique POVsThe single most distinctive claim per videoLLM extracts 1 POV per video. Tagged with that video's SUBS and VIEWS for credibility context.
Sponsored flagYES/NO — explicit sponsorship from a brand related to the queryLLM scans transcript for "sponsored by", "paid partnership", "brought to you by". Only flags if sponsor is relevant to the search query.
Comment Breakdown% split of viewer themes in comments. Hidden if insufficient data.LLM clusters comments into themes with % estimates. Card is only shown if 2+ meaningful themes are found.
Download AnalysisPDF of the full dashboardBrowser print() with print-specific CSS — renders cleanly on white background.
Client IntelligenceWhat the analysis means for a specific client, with recommendationsFrontend sends compiled context back to Modal to query Gemini for bespoke brand strategies.
TagTypeMeaning
HIGHPriorityAct immediately — clear gap or risk identified
MIDPriorityImportant but not urgent — plan into next quarter
LOWPriorityNice to have — revisit when capacity allows
SEffortSmall — days of work
MEffortMedium — weeks of work
HEffortHeavy — months, likely cross-functional
contentCategoryContent strategy or creation
productCategoryProduct or feature related
commsCategoryCommunications, PR, messaging
mediaCategoryMedia, paid, creator partnerships
researchCategoryFurther research or monitoring needed
project/ ├── frontend/ ← Deployed to Vercel (Connected to GitHub) │ ├── index.html ← Main app UI │ ├── style.css ← All styles │ ├── app.js ← Client logic and Modal fetch requests │ ├── docs.html ← This page │ └── report.html ← PDF Export Template │ └── backend/ ← Deployed to Modal via CLI └── modal_app.py ← FastAPI server, Scraping logic, Gemini Prompts

To update the backend logic: modal deploy modal_app.py
To update the frontend UI: Push changes to your GitHub repository and Vercel will rebuild automatically.