Method · 8 minute read · June 2026
Most audience research is vibes. This is a method you can repeat and check. Define the reader, map the channels, keep the active ones, flag the over-performers, then cluster. Seven steps, with the AI builder crowd as the worked example.
Someone says "our people love deep technical content," everyone nods, and three months of work gets pointed at a guess.
I wanted something I could repeat and check, so I built a simple method and ran it on the hardest audience I know: AI engineers and technical leaders shipping agents in production. The output of that run is the eight winning topics I wrote up separately. This post is the method itself. None of it is specific to AI. You can point it at your own audience this afternoon.
Not "developers." Not "technical buyers." I wrote my reader as one sentence: AI engineers and technical leaders building, deploying, and scaling AI agents and LLM applications in production. That precision feels excessive until step six, when it becomes the test that tells you which channels to throw out. A vague reader gives you a vague map.
It is tempting to brainstorm topics. Do not. Start with the places your reader already chooses to spend attention. I found 21 channels and sorted them into six buckets: agents and orchestration, LLM app development, infrastructure and MLOps, coding agents and tooling, research explainers, and builder-facing vendor channels.
They ranged from Gabriel Mongaras at 14.4K subscribers to IBM Technology at 1.71M. Size is not the signal. Fit is. A 30K-subscriber channel that talks only to your reader beats a million-subscriber channel that talks to everyone.
Of the 21 channels, only 10 had posted in the last 90 days. Some of the most influential names, Karpathy, Yannic Kilcher, AI at Meta, had gone quiet. A channel that is not posting cannot tell you what is working now. I set the dormant ones aside and worked from the 10 that were active. You want a snapshot of the present, not the hall of fame.
Each step narrows the field from "channels that might reach my reader" down to a short list of themes I have evidence for.
Source: 90-day public YouTube snapshot, June 8, 2026.
This is the heart of it. For each active channel, take the median views of its last 90 days, then flag every video that beat that median by 2x or more. Ranking by raw views just rediscovers that big channels are big. Ranking by over-performance tells you which topics broke out relative to a channel's own norm.
It also makes small and large channels comparable. A from-scratch agent tutorial did 947 views, but that was 4x its median, so it counts. A speaker-labels demo did 37K, which was 48.6x theirs. Both are real signal. Forty-one videos cleared the bar.
Tag every over-performer with a plain description of what it is about, then let the groups emerge. Mine collapsed into eight: agent harnesses, Claude Code as a platform, how to build agents, reliability and evals, local inference, autonomous data agents, voice, and a brand-new term, OpenClaw. The point of clustering is to stop reacting to individual videos and start seeing the shape. I broke down each theme with its numbers in part one.
This is the step everyone skips, and it is the one that keeps the map honest. A whole pile of channels looked on-target and were not. Cutting them is the work. If your research only adds channels and never rejects any, you are collecting, not researching.
Keep · on the reader
Cut · look-alikes
The method is audience-agnostic. Define your reader in one tight sentence. List the channels that genuinely reach them. Drop the dormant ones. Flag the videos that beat each channel's own median by 2x. Cluster what is left. Be ruthless about look-alikes. What you end up with is a short list of themes you have evidence for, not a brainstorm you have to defend.
One discipline made the whole thing trustworthy: I treated anything I could not verify as not verified. A couple of leads did not resolve cleanly, and I left them flagged rather than guess. I wrote about that, and the findings that surprised me, in the last post.
The whole method in one breath
Specific reader → real channels → only the active ones → rank by over-performance, not raw views → cluster → cut the look-alikes → verify what you can, flag what you cannot.
I am the founder and CEO of Leo AI. I spend most of my time on one question: how do specific audiences find and watch video, and what makes some of it travel. I write about what the data shows. More about Leo.