This lesson explains how a language model reads a page and which parts of its answer an auditor should care about. It introduces the four signals every later lesson relies on, reads a real API response from the demo site, and shows how extraction quietly fails when the input is wrong.
What this lesson covers
- Salience, mentions, type and MID, and what each tells an auditor
- How to interpret salience values, from a dominant subject to a diffuse page
- A live extraction on a real page, including an awkward result
- The difference between versions 1 and 2 of the Natural Language API
- Boilerplate, thin content and rendering as the three common failures
- Four checks that validate a corpus before analysis
Of everything the Natural Language API returns, four fields matter to an auditor. Salience is a share of a page’s attention, so values across a page sum to roughly one; above about 0.3 a page is clearly about an entity, and a page with nothing above 0.12 has no clear subject. Mentions show how often an entity appears and whether as a proper noun. Type drives what a machine expects of the thing. The MID ties a mention to a specific Knowledge Graph node, which is what lets an auditor prove two names mean one thing or catch a link to the wrong thing. Salience is grounded in published research from 2014, which is why the course can defend it in a client meeting.
A real extraction from the demo site shows why reading the output matters: on a page about entity strategy, the top entity is the word “entities”, typed as an organisation. The lesson then covers the version trap — version 2 of the API removes the salience field altogether, so every prominence measure would silently disappear — and explains why the course tools default to v1. It closes with the failure that broke the course’s own analysis: an extraction that captured 121 characters of a 15,399-character article, with no error message anywhere, and the four validation checks that would have caught it.
This lesson continues Module 1, Entity Foundations for Auditors, in The AI-powered Entity Audit, and leads into the 5C Entity Audit Scorecard. It suits SEOs who have seen salience scores in tools without knowing how to interpret them, and anyone about to run an entity extraction on a real site for the first time.