On-Page Entities Past the ~50-Token Mark Probably Add Little Ranking Value

Content briefs built around 100 scraped entities pad the target list with bottom-of-the-barrel terms that probably do very little for rankings.

Kyle Roof reports using two entity libraries — Google's NLP API and public semantic libraries — to extract and evaluate on-page terms.

Google's NLP API only works with seven or so languages, so the public semantic libraries cover the remaining languages.

In Roof's experience, once an extraction passes maybe the 50-term or 50-token mark, a page may not need 100 entities because the low-frequency tail terms at the bottom of the list are probably not doing much to move rankings.

Furthermore, rather than chasing a raw entity count, practitioners might want to score extracted entities by how many times each term appears and build the target list from those frequency counts.

Prioritizing the highest-frequency terms leans into Term Frequency via TF (Term Frequency) scoring, a mechanism that remains alive today despite being dismissed as outdated in the AI era.

Alongside frequency scoring, isolate what the terms are and where they sit on the page while excluding on-page elements that do not drive rankings.