YouTube idea generator: API for the ranking, a model for the angle
Fetch top videos with the YouTube API, score titles and engagement, then ask Claude or ChatGPT for ideas you can actually shoot. Perplexity for the niche, not for the view count.
Use this stack as a page: the YouTube intel generator.
The viral YouTube idea generator is a research pipe: find what already works in a niche, break down titles and keywords, then emit angles that are not a clone of yesterday’s #1.
The card’s pipeline is honest: find trends → analyze → generate ideas. The dishonest part is “viral” as an output type. The API can tell you what did well. It cannot promise the next one will.
YouTube Data API is not a row on our matrix. Do not invent /tools/youtube-api. It is a Google API key, quotas, and a crawl you are allowed to run.
The stack we can name:
- Fetch — YouTube API (your project, your quota).
- Analyze — a spreadsheet or n8n code node: views, velocity, duration, title tokens. Not a model.
- Perplexity — what the niche is about this month, with sources.
- Claude or ChatGPT — titles, angles, keywords you could shoot.
- NotebookLM — optional: ground ideas in your existing scripts so you do not contradict yourself.
- n8n — the Run Automation button.
This is not a ranking-hacking SaaS from us. It is how you stop asking ChatGPT “give me viral ideas” with an empty context window.
Why a chatbot cannot see YouTube
ChatGPT and Claude do not have a live view-count feed. If they list “top videos,” they are remembering, guessing, or browsing in a way you cannot audit. The API returns viewCount, publishedAt, tags (when present), and the title that actually ranked.
Perplexity is the right tool for “why is this niche moving” — news, tools, drama — the same job we give it in the research stack and in Perplexity vs Google. It is the wrong tool for “sort my competitor’s last 50 uploads by views.” That is the API.
Engagement is not just views. A 8-minute video with a high like/view ratio is a different lesson from a Short that got lucky. Your analyze step should split long vs short or you will generate 40-minute ideas from 15-second winners.
The jobs (and what each must not do)
Fetch owns the list
Query by topic, channel ids you care about, or a search string you already use. Cap results. Store video id, title, description snippet, duration, published date, views, comments.
Do not scrape the public website in a gray bot if you have an API. Do not fetch the entire site every ten minutes; quota is a feature.
Analyze owns arithmetic
Normalize by age (views per day) so a three-year-old evergreen does not look like this week’s trend. Extract title bigrams. Flag numbers, questions, “I tried,” year stamps.
A model should not compute views / days. Models drop zeros.
Perplexity owns the world outside the ranking
If the top videos are all about a product launch, Perplexity should say so with links. Claude can then decide whether you have a right to cover it.
Claude owns the idea, not the thumbnail
Claude $20/month paid, long context: paste the top 20 rows plus your channel’s last 10 titles. Ask for:
- Five ideas that overlap the demand without copying a title.
- Keywords (search language, not tag stuffing).
- Format (essay, list, teardown, Short).
- Why this is shootable with your tools — if you are a faceless channel, do not emit “interview the CEO.”
ChatGPT $20/month paid is the swap. Grok is not the YouTube API. Do not mix them.
If the next step is making the video, hand off to the content agent or faceless shorts. This post stops at ideas.
n8n owns the button
n8n ~$22/month cloud. Schedule a weekly fetch. Dump to a sheet. Only then call the model. Humans pick one idea. The generator that auto-publishes from this list is a different, worse product.
One weekly run, six steps
- YouTube API: top videos for the niche + your own channel stats.
- Spreadsheet: velocity, duration split, title tokens.
- Perplexity: three source-backed bullets on niche news.
- Claude: ten ideas, then a human cuts to two.
- Optional NotebookLM: check against your published claims.
- Park the two in the production queue. Do not shoot ten.
Feature-bar translation from the cards: trending topics, high-intent keywords, proven angles, ready-to-use ideas. The first three are analysis. The fourth is a draft until you accept it.
What this stack is not
It is not a view guarantee. YouTube’s ranking is not an AI feature we can mark yes on.
It is not Sora in this post. Thumbnails and generated B-roll wait until you have a title you would put your name on.
It is not stealing a smaller creator’s entire outline. Structure is fair game. Their jokes and footage are not.
The page with the three matrix slots is the YouTube intel generator. The YouTube API stays infra.