This lesson takes one page whose subject a language model misreads and fixes it in two measured rounds, re-running extraction after each. It then turns the cheapest part of the edit into a process that can run across a whole site with a person approving every change.
What this lesson covers
- The subject statement and the four placements
- Round one, measured
- Round two: reducing what competes
- Side effects of a clarity fix
- When the edit will not work
- The same fix across hundreds of pages
The page in the lesson is about entity prioritisation, but the language model reads it as being about “entities”. Round one adds the intended entity to the title, opening, a heading and the meta description and spreads three more mentions through the text — under four per cent more words. Its salience triples, and the verdict is still a mismatch, because the generic word keeps its share. Round two makes that word specific where it stands for the intended entity, and the measurement finally moves: the generic noun falls by more than half and the intended entity rises about four times on the baseline.
The lesson turns that into a checklist, names the three kinds of page the edit cannot fix — pages too short to carry a subject, pages genuinely about two things, and templates — and then moves to scale. On the demo site, 509 pages do not name their dominant entity in the meta description. Rather than writing them by hand or asking a model for generic copy, the lesson builds a brief from audit fields, has Claude draft each description in a fixed format, and keeps a person’s approval between every draft and the live site.
Every number in the lesson comes from live calls to Google’s Natural Language API on the same page text, so you can see exactly how much each sentence moved the result, and the lesson keeps the measurements for the round that did not work as well as the one that did.
This lesson continues Module 5, Remediation: Fixing Entity Representation, in The AI-powered Entity Audit, and is followed by a practical that builds the meta-description process step by step. It suits content and SEO teams who need edits they can measure, not edits they hope worked.