AI SEARCH (GEO)

Extract citable claims from a page

Use when you want to know which sentences on a page an AI assistant could actually quote.

extract-citable-claims.md

Download .md

You are auditing a web page for citability by AI assistants such as Claude, ChatGPT and Perplexity.

Page URL: {{PAGE_URL}}
Page content: {{PAGE_TEXT}}
Primary topic: {{TOPIC}}

A citable claim is a self-contained sentence that states a fact, a definition, a number, a rule or a clear judgment, and that still makes sense when lifted out of the page with no surrounding context.

Return a table with these columns:
| Claim (verbatim) | Type (definition / statistic / process / judgment) | Self-contained? (yes/no) | Why an assistant would or would not quote it | Rewritten version that is quotable |

Then list, as a numbered list, up to eight claims the page implies but never states outright in a liftable sentence.

Constraints:
- Quote claims verbatim in the first column. Do not paraphrase there.
- Do not invent statistics. If a number on the page has no source, mark it "unsourced" in the fourth column.
- Do not suggest adding claims the page has no basis to make.
- Rewrites must be under 40 words, plain language, no em dashes, no "it is not just X, it is Y" constructions.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{PAGE_URL}}
{{PAGE_TEXT}}
{{TOPIC}}

Getting a better result

Paste the rendered text, not the HTML, so navigation and footer copy do not pollute the analysis.
Run it on a competitor page too and compare how many self-contained claims each page has.
Claims that depend on "this" or "as mentioned above" almost never get quoted; fix those first.

Questions about this prompt

When should I extract claims rather than just rewriting the page?

When a page reads well and still never gets quoted. A general rewrite guesses at the problem and changes sentences that were already fine. This works sentence by sentence and tells you which of your existing lines survive being lifted out with no surrounding context, so you edit the dozen that fail rather than the whole page.

What do I need in front of me before running it?

The rendered page text rather than the HTML, so navigation and footer copy do not end up in the claims table, plus the URL and the primary topic. {{TOPIC}} is what stops a stray sentence from an unrelated section being counted as a citable claim on the subject you actually care about.

What comes back, and which part is worth reading first?

A table of verbatim claims with a type, a self-contained verdict, a reason and a rewrite, then up to eight things the page implies but never states in a liftable sentence. That second list is the useful part. Those are positions you already hold and simply never wrote down as a quotable line.

What is the mistake that costs me here?

Pasting every rewrite back into the page in one pass. Any claim marked unsourced in the fourth column should be held, because publishing a tidier version of an unsupported number only makes the unsupported number easier to quote. Fix the claims that depend on this or as mentioned above first.