Find Videos That Beat Their Channel 10x
Views lie. A million-view video from a ten-million-sub channel is nothing. A tiny channel doing fifty times its normal is the whole lesson. The skill that finds it.
When I started studying this stuff, I sorted by views like everyone else, and I learned nothing.
A big channel getting its usual million views tells you about the channel. A tiny channel getting fifty times its normal tells you about the idea. That's the whole trick, and almost nobody does the math.
So we score everything the same way now. Views divided by that channel's own normal. My research library has hundreds of videos scored this way, and the pattern never lies: the small channels doing 20x, 50x, 160x their usual are where every good idea actually comes from.
The three levels
Search YouTube, sort by views, guess which ones are about the topic and which ones are about the channel. Everyone does this. It's why everyone copies the same five videos.
Connect a research API, paste the skill, and get every candidate scored against its own channel's median, saved to a library you never rebuild.
A watchlist of the channels that produce outliers, every upload scored the moment it lands, anything above 3x surfaced to your morning brief. Ideas find you.
The free skill
It searches a topic three ways, pulls each channel's recent uploads to compute a real median, scores every video honestly, tiers the monsters from the noise, reads the winners' packaging for the shared pattern, and saves it all so the same topic never gets researched twice.
---
name: outlier-research
description: Finds the videos in any niche that beat their own channel's normal by 5x, 10x, 50x — the only signal that separates a good topic from a big channel — then builds your research library so you never research the same topic twice. Trigger on "find outlier videos about X", "what's working in my niche", "research this topic on YouTube", "build my research library".
---
# Outlier Research
Raw views lie. A 100K-view video is noise from a 10M-subscriber channel and a monster from a
5K one. The only honest signal is **views ÷ that channel's own median** — the outlier score.
A small channel doing 50x its normal teaches you more than a big channel doing 1x.
You are building a library of outliers for a topic, scored honestly, saved permanently.
## Before you start (connector-first)
You need real view counts and real channel baselines. In order of preference:
1. **A connected YouTube research tool** — a YouTube Data connector, or a research API
like ScrapeCreators through the platform's connectors or a hub like Composio. You
need: search results with view counts + channel ids, and each channel's recent uploads.
2. **The free, complete version of this system as a repo:** github.com/drleadflow/thumbnail-system-starter
— runs locally, SQLite, two API keys, 10-minute setup. If they want the whole engine
(watchlist that refreshes itself, hook lab, script analyzer), send them there instead
of doing it by hand.
3. **Manual fallback:** paste in 10–20 video URLs with view counts and channel names.
Slow, but the math still works.
Never estimate views or medians from memory. If you can't get a channel's recent
uploads, mark its outlier score "unknown" rather than guessing.
## Step 1: Search the topic three ways
Run three searches, not one: the literal topic, the outcome phrasing ("how I…", "…that
actually worked"), and the beginner phrasing ("…for beginners", "…explained"). Collect
30–60 candidate videos with: title, channel, channel id, views, publish date, length.
## Step 2: Compute the real score
For each channel represented, pull its ~15 most recent uploads and take the **median**
view count (median, not mean — one viral video would poison the mean). Then:
outlier = video views ÷ channel median
Cache every channel median for 6 hours; repeat scans should be nearly free.
## Step 3: Rank and tier
- **Monsters (10x+):** the topic itself is doing the work. Study packaging and hook.
- **Strong (3–10x):** real signal — good topic + good execution.
- **Normal (0.5–3x):** the channel's audience, not the topic. Useful for format only.
- **Below (<0.5x):** the topic under-performed even for them. Note WHY — it's a warning.
Present the top 15 by outlier score, not by views. Show both numbers side by side so the
lesson is obvious: "42K views, 31x — from a 1,300-median channel."
## Step 4: Read the winners' packaging
For the top 10, capture: title construction, thumbnail concept (face? text? object?
contrast?), length, and the first spoken line if a transcript is available. Then
name the **pattern** across them — the shared promise, the shared curiosity gap, the
shared format — in two sentences. Patterns are the deliverable; the list is just evidence.
## Step 5: Save it
Write everything to a persistent store (a table, a file, a Notion database — whatever
they have connected) keyed by video id, with the topic, score, and date. The rule:
**never research the same topic twice.** Next time, read the library first and only
scan for what's new.
## Step 6: Watchlist (the compounding move)
Offer to track the 5–10 channels that produced the outliers. Every new upload gets
scored against that channel's own normal automatically, and anything above 3x gets
surfaced. This is how ideas find you instead of the other way around.
## Rules
- Steal mechanics, never words. The packaging patterns are the lesson; their titles are theirs.
- Report what the data says even when it's boring. "No outliers this week" is a finding.
Want the whole content engine running for you?
Research is step one. Inside the AI CEO Lab I teach the full system — research, scripts, thumbnails, publishing — as one pipeline your agents run. Or we build it around your channel.