AI Overview fact checking prompt with a claim by claim evidence ledger
Paste an AI Overview and its cited pages. This prompt separates every claim, checks what each source supports and flags missing context.
- Works in
- ChatGPT, Claude, Gemini
- You need
- The complete AI Overview text · The query that produced it · The text from every cited source · Your country and the date of the search
- Written for
- ai overview fact checking prompt
An AI answer can contain nine sound claims and one decisive error. A single score hides that shape. This prompt keeps every claim separate and forces the citation to earn its place.
What the verdicts mean
Supported means the supplied passage establishes the complete claim. Partly supported means the central idea appears but a limit was lost. Unsupported means no supplied passage proves it. Contradicted means the evidence says the opposite. Not checkable means you need another source.
Use it with the full checking method in are AI Overviews accurate. The prompt creates the ledger. You make the final judgement.
Act as a careful fact checker. I will give you a search query, an AI Overview
and the text of its cited sources. Do not use memory to fill a gap. Judge only
what the supplied sources establish.
First split the overview into atomic claims. One row must contain only one
claim that can be supported or rejected independently.
For each claim, report:
Claim | Importance | Source offered | Exact supporting passage | Verdict |
Missing condition | Freshness risk | Better source needed
Use only these verdicts: supported, partly supported, unsupported,
contradicted, or not checkable from supplied material.
A page about the same topic is not evidence. The exact passage must support
the complete claim. If the overview removes a population, date, geography,
billing period, denominator or exception, mark it partly supported and name
what was lost.
Mark importance as low, medium or high. High means that being wrong could
change a health, money, legal, safety or purchasing decision.
After the table, provide:
1. The three claims most in need of manual checking.
2. A list of cited pages that were topically relevant but did not prove the
sentence attached to them.
3. The primary source type that would settle every unresolved high importance
claim. Name the type, not a source you have not opened.
4. A corrected summary using only supported material. Preserve uncertainty.
Do not give the overview one overall accuracy percentage. The sample is too
small and claim importance is unequal.
Query: [SEARCH QUERY]
Country and search date: [COUNTRY AND DATE]
AI Overview: [PASTE COMPLETE OVERVIEW]
Cited source text: [PASTE EACH URL FOLLOWED BY ITS RELEVANT PAGE TEXT]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[SEARCH QUERY]- The exact wording matters because the answer may be reasonable for one interpretation and wrong for another.
[COUNTRY AND DATE]- Location and time settle whether current rules, prices and availability were evaluated in the right context.
[PASTE COMPLETE OVERVIEW]- Include the full answer rather than the sentence that looks suspicious so qualifications elsewhere are not lost.
[PASTE EACH URL FOLLOWED BY ITS RELEVANT PAGE TEXT]- Paste source text because a model may not be able to open every page and should never pretend that it did.
How to run it
- 01Capture the result
Save the query, answer, citations, date and country while the result is visible. Generated results can change between searches.
- 02Open every important citation
Copy the passage around the claimed fact. Include the heading and publication date so context survives.
- 03Run the claim ledger
Paste the complete material and do not let the model browse silently. The point is to audit a fixed evidence set.
- 04Check the high importance rows yourself
Open the primary source and confirm the passage. The prompt organizes verification but does not replace it.
Questions people ask
Can this prompt tell whether an AI Overview is accurate?
It can show whether the sources you supplied support each claim. It cannot prove that the supplied sources are themselves true or complete, so important claims still need a primary source and human review.
Why does it avoid one accuracy score?
Ten harmless facts and one dangerous error should not become a reassuring ninety percent. A claim ledger preserves both the kind of failure and its importance.
Unsubscribe in one click. We never pass your address on.