What do I do when a paper's cited source does not support its claim?

The discovery is unsettling the first time. You follow a citation to check a detail, and the cited paper does not contain the claim, or contains something narrower, or is about a different population entirely. It is not rare, and finding it means you are doing something most readers do not.

There are three distinct causes and they call for different responses. Sometimes the claim is there but in a section you did not check, so read the whole paper before concluding anything. Sometimes the citing author generalized: the source reported an effect in one setting and the citing paper stated it without conditions. And sometimes the citation is simply wrong, pointing at the wrong work entirely, which happens through reference-manager errors more often than through carelessness.

The generalization case is the most consequential, because it is invisible to anyone who does not check. A conditional finding becomes general, gets cited again in its general form, and after four hops the condition has disappeared from the literature while the citation still points, formally, at real evidence.

When you find one, the useful move is to trace rather than to discard. The chain usually leads somewhere, and knowing where a widely repeated claim actually comes from is a genuine contribution you can put in a footnote or a paragraph.

Check the cited work itself, not the citing paper's summary of it and not the reference list. Penders points out that a citation marks where an idea, method, or critique came from, and that reading the cited work is what lets you judge whether the citation fits (Penders, 2018, p. 2). Compare the exact proposition in the citing sentence against the source's research question, sample, methods, results, and stated limitations, then record whether the source supports the claim, supports only a narrower version, reports a competing result, or says nothing relevant at all.

How often this happens has been measured. Jergas and Baethge define a major quotation error as one where the cited source is not at all coherent with the claim attached to it, which is a different problem from a wrong page number or a garbled reference (Jergas and Baethge, 2015, p. 2). Across the medical studies they reviewed, major errors ranged from 2.2% to 55.0% of checked citations, with a median of 11.5% across 27 studies, and total errors ranged from 6.7% to 83.0%, with a median of 22.5% across 28 studies (Jergas and Baethge, 2015, p. 5). They also warn that the variation between studies is large enough that no single figure transfers to your field (Jergas and Baethge, 2015, pp. 16-20). Treat the numbers as a reason to check, not as an expected error rate.

How far back through a citation chain should I go?

Until you reach a paper reporting original evidence rather than citing someone else. Chains of three or four are common and chains of six exist. Stop when you find data, or when you find that the earliest link contains no evidence either, which happens more often than you would expect.

Keep a note of the chain as you go, with each paper and what it actually said. If the claim matters to your argument, that note becomes a paragraph. If it does not, the note stops you repeating the trace in six months when you have forgotten doing it.

Watch for the loop, where two papers cite each other and no original evidence exists anywhere. It sounds unlikely and it is documented in several fields, usually around a number everybody quotes.

Note which studies the chain routed around, not only where it ends. Greenberg separates three distortions in a citation network, citation bias, where supporting papers are cited and critical ones are not, amplification, and invention (Greenberg, 2009, p. 8). When he corrected for bias against critical content in the network he traced, five of six primary-data papers that had been cited rarely turned out to carry authority (Greenberg, 2009, p. 4). Repeated citation is not independent confirmation, so a chain that ends in real data can still be misleading if the contrary studies were dropped along the way.

Can I still use the claim if the chain breaks?

Only by changing what you assert. You can say the claim is widely repeated in the literature and cite the repetitions, which is a different and often more interesting statement. You cannot present it as established, because you have just discovered that it is not.

The reframing is genuinely useful. "This figure is widely cited but traces to a single unpublished estimate" is the kind of sentence that makes a literature review worth reading, and in some cases it is a finding in its own right.

If you replace the source, pick the replacement for the job the sentence is doing. Penders recommends reviews for broad context in an introduction and empirical papers when you describe a specific contribution in a discussion, and advises prioritizing peer-reviewed work over editorials and preprints, deprioritizing self-citations, and limiting how many citations you stack behind one claim (Penders, 2018, p. 4). A review can define a field-level problem, but a specific number belongs to the study that produced it. Where the mismatch is arguable rather than clear, quote the passage and set out both readings, since Greenberg notes that readers can reasonably disagree about what a text asserts (Greenberg, 2009, p. 8).

If the claim was load-bearing for your argument, treat its collapse as important information rather than an inconvenience. Better to find out now than to have an examiner find it in your chapter.