How to Find a Source That Actually Supports Your Paragraph

Work backwards from the sentence you have already written. Pull out the claim it actually makes, restate that claim in the vocabulary the literature uses, search a database that indexes that literature, then open the retrieved text and check whether the passage supports your specific sentence. Most weak citations in student work are not invented references. They are real publications attached to a sentence they do not actually back up.

Start from the claim, not the topic

A topic search gives you a reading list. A claim search gives you a citation. The difference shows up in what you type into the box.

Before: your paragraph argues that hybrid work lowers burnout among office employees. You search "remote work", open the first review that looks relevant, and cite it.

After: you underline the claim itself, hybrid arrangements reduce self-reported burnout among knowledge workers. That sentence contains a population (knowledge workers), an intervention (hybrid arrangement), an outcome (self-reported burnout) and a direction (a reduction). Each of those is a search term, and each is something the source has to contain before the citation holds.

The test behind this is older than search engines. Writing about how references function in scholarly work, Sellitto relays Metcalfe's proposal that a literature review can be appraised by treating any citation as a form of argumentative evidence, in much the same way that evidence is considered in a judicial process (p. 6). Evidence in that sense is always evidence for something. A source that is merely about your topic is not evidence for your sentence.

Turn the claim into a query, then run it twice

Corpus linguists hit the same problem you do, only they measure it. Kljajevic and Šarić, comparing ways of retrieving proverbs from Croatian and Norwegian language corpora, describe the trade-off plainly: you can use restrictive queries that mostly retrieve relevant results but exclude some relevant examples, or broad queries that return a higher number of hits while also including many non-relevant ones (p. 139). Their material was proverbs rather than thesis literature, but the shape of the problem carries over. Too narrow a query drops sources you needed, and too broad a one leaves them buried under results that have nothing to do with your claim.

So run both. Start narrow, with the exact construct from your claim ("hybrid work" AND burnout AND "knowledge workers"). If you get under ten hits, widen one element at a time: swap burnout for exhaustion, drop the population, allow "flexible work arrangements". Iterating is the normal case, not a sign you searched badly. In their survey of large language models, Zhao and colleagues describe retrieval pipelines that rewrite or expand the input query and run multi-turn refinement when a single round of retrieval proves insufficient (p. 2010). That is machine retrieval rather than library search, but if the systems built to do this automatically need several passes, your first query was never going to be the winner either.

Where you search matters as much as what you type. Gusenbauer, comparing the sizes of 12 academic search engines and bibliographic databases, states the constraint directly: queries define the line between what data can and what data cannot be retrieved by the regular user (p. 184). He also notes that Google Scholar became the number one go-to information source in academia (p. 178), which is exactly why leaning on it alone is risky. Two further points from the same study are worth keeping in mind when you judge your results: search systems often contain a significant portion of duplicates and cataloguing errors that inflate their apparent size without giving you anything new (p. 184), and when relevance is accounted for, a larger scope yields better results than a smaller one (p. 178). Ten hits on your screen are not ten usable sources, and one database is not the literature.

If you would rather skip the query-tuning entirely, this is the job cytado's source finder does. You paste the sentence, it searches the corpus and returns publications with the passage and the page. The academic sources under this article were found that way.

Read the passage, not the abstract

This is the step that separates a citation that survives review from one that does not. An abstract tells you what a paper is about; only the passage tells you whether it says your sentence.

The scale of the mismatch problem is documented. Safran and Çalı, reviewing citation reliability in AI-assisted writing, report an analysis of AI-generated medical content in which 47% of references were fabricated, 46% were authentic but inaccurate, and only 7% were both authentic and accurate (p. 700). That middle number is the one relevant here. Nearly half of those references pointed at something real while getting the details wrong, which is precisely the failure that survives a quick existence check and then collapses when your supervisor opens the paper. The same authors recommend building DOI verification through Crossref or PubMed APIs into editorial workflows (p. 701), which is the check cytado's bibliography checker runs across a whole reference list at once.

Mismatch has a vocabulary in the literature on citation practice. Subramanian and colleagues identify misquoting or misrepresenting sources as the most common citation problem, one that spreads inaccurate information, and their catalogue of poor practice includes misinterpretation of results, unnecessary extrapolation of the outcomes of cited work, and ignoring a more suitable reference (p. 99). "Unnecessary extrapolation" is the diagnosis for most student citation errors: the source is real, it was read honestly, and the sentence built on it claims more than the study measured.

Before: "Hybrid work eliminates burnout in office teams (Kowalski, 2022)."

After: "In a survey of 380 employees at three IT companies, hybrid work was associated with a moderate decline in self-reported exhaustion (Kowalski, 2022, p. 47)."

The second version is narrower, and it is the one that holds when someone opens page 47. If you cannot narrow your sentence to what the source actually measured, you have the wrong source rather than a phrasing problem.

Note the page number while the text is open

Take the page while you are reading the passage, not at the end of the week when you have thirty tabs and no memory of which one carried the finding. Two habits save the most time here. Record the page from the printed page number visible in the PDF rather than from the viewer's page counter, and record a DOI rather than a URL.

The second habit has hard evidence behind it. Sellitto examined 1,068 web-located citations across 123 conference papers and found that some 46% of them could not be accessed, with the HTTP 404 message accounting for 61.5% of the failures; the missing citations amounted to 22% of all citations in the sample (p. 2). Those were published, peer-reviewed papers. A URL you paste into a bibliography today has the same decay ahead of it, while a DOI resolves to the publisher's current address for the article.

If you want the wider view of where to look before you get to the paragraph level, our guide on finding reliable thesis sources covers database choice and credibility screening.

The short version

1. Underline the claim in your paragraph and name its parts: population, intervention, outcome, direction.
2. Translate the claim into the vocabulary of the literature, not your own phrasing.
3. Run a narrow query, then a broad one, in at least two databases.
4. Open the retrieved text and find the passage that carries your claim.
5. Narrow your sentence to what that passage supports, and record the page and DOI while you are there.