How do I check that every claim in my thesis is supported by its citation?

Most people check citations by rereading their chapters, which is the least effective way to do it. Reading your own prose activates the memory of writing it, and the claim feels supported because you remember believing it. Verification needs the opposite: the claim in isolation, next to the evidence, with no surrounding argument to make it sound reasonable.

So separate the claim from the chapter. Extract every sentence that carries a citation into a list. For a thesis this is somewhere between two and six hundred rows, which sounds crushing and takes less time than expected, because most rows resolve in under a minute once the source is open.

Work by source rather than by chapter. Open one paper, check every claim in the thesis that cites it, then move on. That saves the opening and reorienting cost, which is most of the time in this task.

Give the pass a fixed budget. Three days near the end of writing, done properly, is enough for a thesis, and it is one of the highest-return uses of three days available to you. Examiners check citations, and they usually start with the ones that sounded surprising.

The error rates justify the time. A meta-analysis of quotation accuracy in medical journal articles found pooled rates of 11.9 percent major errors, 11.5 percent minor errors, and 25.4 percent total errors, where a major error means the cited source is not coherent with the claim the citing author attached to it (Jergas and Baethge, 2015, p. 2). Those pooled figures come from roughly 27 studies of medical articles with very high heterogeneity between them, so treat them as a reason to check rather than as a rate for your own field (Jergas and Baethge, 2015, p. 11). The authors do report that the finding of substantial quotation inaccuracy held up across their sensitivity and subgroup analyses (Jergas and Baethge, 2015, p. 16).

What does a claim-level audit look like in practice?

One row per cited sentence: the sentence, the source, the page, and a verdict of supported, partly supported, or not found. Partly supported is the important column, since it catches claims that are broader than the evidence, which is the most common failure and the hardest to see while writing.

Decide what each row means before you write the verdict. Compare claim and source on five points: meaning, scope, population, method, and result. Ask whether the source makes the claim at all, whether it covers the same participants or texts or period, whether you turned a qualified result into an absolute one, whether you read an author's interpretation as an observed finding, and whether the number and its direction and uncertainty came across intact. A reference that carries an argument needs enough context to show its relevance and whether the cited findings support or oppose your conclusion, which is a different job from a reference to a dataset or a method (Penders, 2018, p. 3).

Watch for the failure where the source exists and says something close to your sentence, but not the same thing. Greenberg calls this citation diversion, and the network analysis shows how far it travels. One primary paper found no beta-amyloid precursor protein in three of five patients and only a few fibres in the other two, yet later papers described that evidence as confirming a broad accumulation claim, and the altered version went on to collect 7,848 supportive citation paths over the following decade (Greenberg, 2009, p. 3). Your verdict column should record the strength of support, not only its presence.

Do not fix as you go. Marking and moving keeps the pass fast and keeps you in verification mode rather than drifting into rewriting, which is a different task with a different rhythm. Fix everything in one editing pass afterwards.

Sort the finished list by verdict. The not-found rows need a source or need cutting. The partly-supported rows need the sentence narrowed, which is usually a change of three or four words.

Can an abstract support a claim on its own?

Rarely. Abstracts drop the conditions that make findings true, so a claim verified against an abstract is often broader than the paper supports. Use abstracts to decide whether to check the full text, never as the evidence you verified against.

The specific things abstracts lose are population limits, effect sizes, the analysis actually run, and the authors' own caveats. Any of those can be the difference between your sentence being true and being an overstatement.

The same network analysis gives a concrete case. Seven papers made 17 citations to 12 abstracts whose content was misrepresented, and a claim about one form of beta-amyloid being more common than another was absorbed as established fact even though the cited item was a conference abstract rather than a full paper (Greenberg, 2009, p. 5). Penders makes the related point that citation counts and database entries are not evidence of quality and do not tell you whether a citing work endorsed or challenged the claim, so consulting the publication's actual content is the only reliable check (Penders, 2018, p. 2).

There is one exception worth naming. When you are citing a paper for its existence or its topic rather than its result, the abstract is sufficient, because that is all your claim requires.

Whichever route you take, record the verification. A column noting the date you checked each claim turns a nervous memory into something you can point at when an examiner asks how carefully the references were handled.