Start from one claim, not the whole answer
Pick the single sentence that would change your reading of the paper if it were wrong: a number, a comparison, a claim about what the authors concluded. Broad summaries are hard to verify and rarely worth it. A specific claim can be checked in under a minute.
Read the quote before you read the answer again
Open Where this comes from and read the quoted lines on their own. Three questions settle most cases:
- Does the quote actually contain the fact the sentence asserts, or only the topic of it?
- Does the quote carry a qualification the answer dropped — a subgroup, a time period, an "in this sample"?
- Is the quote from a part of the paper that can support the claim at all? A sentence from the introduction is often the authors describing someone else's work.
A quote that is merely on the same subject is not support. That gap between "related text" and "the text that establishes this" is where most confident-sounding errors live.
Go to the page and read around it
Open your PDF at the page shown and read the paragraph the quote sits in, plus the sentence before and after. Tables and figures matter here: a number in the body text is often a rounded or partial version of the value in the table, and the table caption usually tells you the denominator. If the claim is quantitative, put the paper's value and the answer's value side by side and compare units, sample and time period, not just the digits.
When there is no quote
Sometimes the section reads No verified quotes from the document are available for this result. That is an honest report, not a verdict on the answer. It usually means one of three things:
- the answer is a synthesis or a piece of writing rather than a set of factual claims, so there is no single line to point at;
- the model paraphrased instead of quoting, and nothing it offered matched the document word for word; or
- the PDF's text layer is poor — a scan without OCR, or a layout that extracts as scrambled text.
In that case, treat the answer as a starting point and verify the claims yourself with the checklist in How to check an AI explanation of a research paper. If a whole document extracts badly, the quickest fix is usually a text-based copy of the same paper rather than a scan.
Keep a short record
For each claim you rely on, write down:
- Claim: the sentence as you would use it.
- Quote and page: the lines you actually read, and where they are.
- Fit: does the quote establish the claim, only relate to it, or contradict it?
- Status: verified, qualified, or not verified.
Two habits make the difference over a long reading session: never carry a number forward that you have not seen in the paper, and never repair a mismatch by guessing. Mark it unresolved and move on. For separating what a study found from what its design supports, see How to read a paper's limitations. If a reference in an answer points outside your document, see How to spot a fabricated citation.