This lesson adds a language model to the audit for the questions the NLP API cannot answer, and shows how to make its output auditable rather than merely readable. It is as much about validation as capability: every LLM finding is checked before it is used.
What this lesson covers
- Where entity extraction stops and an LLM adds something
- Using a JSON schema with Claude’s structured outputs
- Reading a real response and the attribute gaps it surfaces
- Recognising plausible, invented entities
- A three-check validation process and confidence tiers
- Choosing between the NLP API and an LLM for each audit question
An NLP API returns what is literally on a page. It cannot say what a page is implicitly about when the subject is never named, how two entities relate to each other, or which attributes a reader would expect that the page never covers — and that last one is a genuine content gap with no query volume and nothing to crawl. The lesson shows how to ask Claude for all three against a JSON schema, so the response has the same shape on every page and every run, including a confidence score and an explicit list of what the model is unsure about.
The second half covers the part vendor demos leave out. Language models occasionally return a plausible, on-topic entity that is not in the text, with the same confidence as the real ones, and once it reaches a content brief nobody can say where it came from. Three cheap checks catch most of these — whether an explicit entity appears verbatim, whether the NLP API agrees, and whether the Knowledge Graph knows it — and findings are tiered from solid to candidate rather than simply kept or discarded. The lesson ends with a decision table for which engine answers which question.
This lesson continues Module 2, Diagnostics: Extracting the Truth About Your Site, in The AI-powered Entity Audit, and is followed by a practical that runs the Claude analysis notebook step by step. It suits SEOs who want to use LLMs in audits without putting unverified findings in front of clients.