<b>Entity salience beats keyword density — but by how much?</b>

<b>Entity salience beats keyword density — but by how much?</b>

Keyword density died years ago; the replacement question is whether <i>entity salience</i> (how central a named entity is to the document, per NLP APIs) predicts ranking better than term frequency.

I ran 200 top-ranking pages through an entity-extraction API, recording salience scores for the primary entity, then correlated against position. I did the same for plain keyword density.

— Keyword density vs position: correlation ~0.04. Effectively noise. As expected.
— Primary-entity salience vs position: ~0.21. Weak, but real and consistent across niches.
— The stronger signal was <i>entity completeness</i>: pages mentioning the cluster of co-occurring entities that top pages share (the "expected entities") ranked notably better than pages with high salience but a thin entity set.

The lesson: don't optimize how often you say the keyword; optimize whether you mention the things a knowledgeable document <i>would</i> mention. Salience matters; co-occurring-entity coverage matters more.

A caveat: correlation isn't causation. Top pages may cover more entities <i>because</i> they're better, not rank because of it.

Method note: salience via a commercial NLP API; expected-entity set built from intersection of top-3 results.

Confidence: medium.