Your winner is a denominator artifact

We needed to choose a content lane, and we wanted the decision made by data rather than taste. So we built a miner: 24 search queries produced 180 candidate channels, 108 of those were profiled in depth, and for each we pulled the 30 most recent long-form videos. That gave us 2,916 videos to score.

The first metric was the obvious one — views divided by subscriber count. A video that massively outperforms its channel's size is evidence that the topic, not the audience, did the work.

The raw counts were clear. One category had 21 videos above the 10× threshold. The runner-up had 10. On the numbers it wasn't close.

The numbers were wrong.

Eighteen of twenty-one were the same channel

When we grouped the winning videos by source rather than counting them, 18 of the 21 came from a single channel with 656 subscribers — and its actual subject matter wasn't even the category we'd assigned it to.

That's the denominator doing all the work. Divide by 656 and almost anything clears 10×. We hadn't discovered a hot topic; we'd discovered one small channel, twenty-one times, and let it vote twenty-one times.

Counting rows treats every row as an independent observation. When your rows share a source, they aren't independent, and your count is measuring the source.

So we added a second metric

The fix for a ratio that flatters small denominators is a ratio that doesn't use them. Our second metric was views divided by that channel's own median views — how far a video outperformed its own channel's baseline, regardless of size.

That metric produced a different leader. It also produced a different failure.

Eight of the top ten were a single non-profit that averages around 200 views per video and had once landed a 2.1-million-view piece. Divide 2.1 million by a median of 200 and the ratio is enormous. But nothing about that result was repeatable — it was one institutional fluke, amplified by a tiny median.

Each metric fails in the opposite direction

• Views ÷ subscribers flatters tiny channels. Any channel small enough will produce huge ratios on ordinary videos.
• Views ÷ own median flatters dormant channels. Any channel quiet enough will produce huge ratios on a single fluke.
• Run both, and demand a candidate win on both. The failure modes don't overlap, so surviving both is meaningfully harder than surviving either.

The category we ultimately chose won on both metrics with 31 distinct channels clearing the bar — no single source dominating, no institutional outlier carrying the average. That distinctness count mattered more than the raw score.

There was a third contaminant, and it wasn't statistical

A third category scored well on both metrics and was still wrong. Its high performers turned out to include a hardware review channel and a professional practice — both pulled in by a query whose wording matched their content, neither remotely comparable to what we could produce.

No amount of metric tuning catches that. It required opening the actual channels and looking. Automated scoring tells you which rows are extreme; only inspection tells you whether the rows belong in the table at all.

The general form

This is not really about video. It's about every ranked list a business makes decisions from — best-performing campaigns, highest-converting pages, most valuable customer segments.

Any time your metric is a ratio, ask what happens when the denominator gets small. Any time you rank by count, ask how many distinct sources those counts came from. And before acting on the leader, open it and look at it, because a number can be correct and still not mean what you think.

Either metric alone would have sent us down a road that wasted months. Running both, and then checking by hand, cost an afternoon.