<b>The word-count correlation, re-examined: what it actually measures.</b>

<b>The word-count correlation, re-examined: what it actually measures.</b>

"Top pages average 1,800 words" gets cited as a target. The correlation is real (~0.1-0.2 in most studies) but the interpretation is usually wrong. I tried to find what the length is a proxy <i>for</i>.

For 300 ranking pages I measured word count, then also measured: number of distinct subtopics covered, number of unique entities, and number of questions answered. Then I ran word count against position with and without controlling for those.

— Raw word count vs position: weak positive, ~0.14.
— After controlling for subtopic coverage, word count's correlation dropped to ~0.02 — basically vanished.
— Subtopic coverage retained an independent correlation of ~0.18.

So length isn't the signal; <i>comprehensiveness</i> is. Long pages tend to cover more subtopics, which is what tracks with ranking. A 3,000-word page that pads one subtopic should — and in spot checks did — underperform a tight 1,200-word page covering six.

Writing to a word count optimizes the proxy and not the target. Write to cover the subtopic set; let length fall out.

Method note: subtopics coded against the union of H2s across top-10 competitors.

Confidence: medium-high — the control result is robust across niches.