How to find a source for a claim you already wrote
A sentence is the worst possible keyword query and the best possible semantic one, which is why the first move is to paste the whole thing into OpenAlex or Semantic Scholar rather than picking four words out of it. Read three abstracts, take the field's vocabulary, then go back to quoted phrases. And decide before you start that you are willing to change the sentence. A search that ends with you rewriting the claim to match the evidence is a successful search, not a failed one.
Writing first and sourcing afterwards is not shameful. Almost everyone does it under deadline. The risk is that you begin with a conclusion and go looking for permission, and that is how people end up attaching a real paper to a claim it does not support. Go in prepared to lose the argument and this works fine.
First, does it need a source?
Some sentences do not. Your own reasoning about your own evidence needs no citation. Standard mathematics needs none. Common knowledge needs none, which the Harvard Guide to Using Sources defines as information generally known to an educated reader, while advising you to cite if you are unsure. Ideas and interpretations rarely qualify.
What always needs one: a specific number, a research finding, a claim about what studies show, a historical particular, anyone else's idea or phrasing. Read your paragraph and mark which sentences are which. People often discover that the sentence causing them anxiety is their own analysis, which needs nothing, while the unremarkable statistic two lines up needs a source and does not have one.
Turn the sentence into a query
Strip it to the claim
Take out the hedges, the transitions and the rhetoric until only the assertion remains. "It is widely recognised that excessive social media use among adolescents can contribute to declining academic performance" reduces to: social media use, adolescents, academic performance, negative association. Four elements. That is what you are searching for.
Translate into the field's vocabulary
Researchers rarely say "traffic getting better." They say travel time reliability, vehicle kilometres travelled, mode share. Your search fails because you are using ordinary words in a specialist index. The quickest fix is to open one paper on the topic and steal the language from its abstract and keyword list.
Search Google Scholar with quoted phrases
Put distinctive multi-word terms in quotation marks so they are matched as phrases. Use the year filter on the left if your claim is about recent evidence. The allintitle: operator restricts results to papers whose titles contain your terms, which is a fast way to find work squarely about your topic rather than work that mentions it. University library guides such as UTSA's Google Scholar guide list the rest of the operators.
Use semantic search when you can't guess the words
This is the step most students have never tried, and it is the one built for this exact job. OpenAlex's semantic search returns works whose meaning is closest to your query even when the wording differs, and it explicitly works better with longer input: paste in your whole sentence or paragraph rather than keywords. Semantic Scholar does something similar across a very large index. When you do not know the technical term, meaning-based search finds it for you, and then you can go back to keyword search armed with the right words.
An opinion, since most guides put this step last if they mention it at all. For this particular job, starting from a sentence you have already written, semantic search should be your first move and keyword search your second. The reason is in the shape of the input. You are not starting with a topic, you are starting with a sentence, and a sentence is the worst possible keyword query and the best possible semantic one: every hedge and connective that ruins a keyword match is signal to a system matching on meaning. Keyword search is still the sharper tool, but it only becomes sharp once you know the field's words, and the first search is exactly when you do not. Run the sentence through semantic search, read three abstracts, steal the vocabulary, then go back to quoted phrases.
Find a review, then mine its references
Search your topic plus "systematic review" or "meta-analysis." A good recent review is a curated map of the primary literature: its text tells you which finding belongs to which study, and its reference list hands you the citations. Take the primary study, not the review, when you are citing a specific finding, and open the primary study before you cite it. Passing along a claim you have only seen quoted is how errors travel through a field for decades.
Walk the citation graph
Once you have one relevant paper, you have two directions. Backwards: its reference list, for what it built on. Forwards: Google Scholar's "Cited by" link, for what came after, including anything that contradicted it. scite goes further by classifying citing statements as supporting or contrasting, which tells you quickly whether a finding held up.
Ask the indexes yourself
Every search box above is a website wrapped around a free API, and typing the API into your browser's address bar is quicker, more precise, and free of whatever ranking the site applies on the way out. These are the three we query in production, so the URL shapes below are the ones that work rather than the ones the documentation implies.
CrossRef will take a whole reference string. Put https://api.crossref.org/works?query.bibliographic= in front of your reference, or of whatever fragment you have, and add &rows=5 to keep the answer readable. This is the search for a half-remembered citation rather than a topic: paste the authors, the year and as much of the title as you can recall in one string and let the matcher do the work. Add &mailto= and your own address and CrossRef moves you into its faster pool for polite users.
OpenAlex will take a topic and a date range. https://api.openalex.org/works?search= plus your terms, then &filter=from_publication_date:2020-01-01,to_publication_date:2024-12-31 to bound the period. The answer opens with a meta.count, which is how many works in the whole corpus match your phrasing, and that one number tells you whether your query is too narrow or hopelessly broad before you have read a single result. One change worth knowing about: since 13 February 2026 OpenAlex has asked for an API key and retired the old email parameter. A key is free from your account settings and raises your daily allowance substantially. Without one you still get a small anonymous budget, and when it runs out the answer is an HTTP 429 that tells you plainly what happened and when it resets, rather than silently returning nothing.
Unpaywall will tell you where the free copy is. https://api.unpaywall.org/v2/ plus the DOI plus ?email= and your own address returns every legally deposited free copy of that paper, anywhere, plus an oa_status naming which flavour of open access it is: gold, hybrid, bronze, green or closed. The email is not optional, and leaving it off returns a 422 that says so. Which of those copies to actually click is the next section.
One warning that cost us a day. Every one of these indexes lets you sort by citation count, and it is the most tempting button on the page, because a highly cited paper feels like a safe citation. Do not press it. Sorting by citations ranks the whole corpus that matched your words, not the results you were shown, so the winner is whatever famous paper happens to contain one of your terms. We measured this on OpenAlex across twenty real queries on 11 August 2026 and an off-topic paper took first place on all twenty: a query about class-size reduction in schools was topped by a heavily cited paper from an entirely different field that shared a single common word. Rank by relevance, take the first page, and judge influence yourself from the citation counts shown beside each result.
Getting past the paywall
Your library first, because that is what the subscriptions are for. After that, one rule carries most of the value: Unpaywall hands you a list of free copies and nominates one as best, and you should ignore the nomination. The nominated copy is frequently the publisher's own, which is precisely the one most likely to refuse the request. Work down the list yourself, repository copies first. We walk through the full ladder, including the arXiv and bioRxiv URL tricks and what to do when a publisher page blocks you outright, in how to find peer-reviewed articles.
Whichever route you take, get the whole paper. An abstract is not enough to cite from, because the abstract is where the qualifiers get dropped. That is the same mechanism that makes any summary overstate a finding, and it applies just as much to the authors' own summary of their own work.
Then read it before you cite it
Finding a plausible paper is the easy half. Confirm that it actually supports your sentence: check the sample it studied, check whether it reports an association or a cause, and find the specific line that backs your claim. Copy that line into a note next to the citation. This is the check that catches the failure nobody talks about, where the source is real but says something else, and it takes about ninety seconds.
While you are there, confirm the paper is what it appears to be. A quick look at the journal, and a search of the title in a retraction database, covers you against the ways an article can be less real than it looks.
Then turn the source into a reference without retyping anything. Paste the DOI into our free citation generator and it pulls the authors, year, journal, volume and pages from the publisher's own registered record and formats them in APA, MLA, Chicago, Harvard, IEEE, Vancouver or numbered. Nothing in that entry is composed, which is the point: hand-typing a reference off a PDF's title page is where wrong volume numbers come from, and a wrong volume number is the detail that makes a marker check the rest of the list. If you only have the identifier and want to see what it is registered to before you commit, the DOI lookup shows you the record on its own. Both are free and need no account.
When nothing supports the claim
You have searched properly and found nothing. This happens, and the honest responses are all better than the alternative. In order of preference:
Narrow it. Move the sentence to meet the evidence. "In cities that introduced a cordon charge, traffic entering the zone fell in the first year in several before-and-after studies" is defensible where the sweeping version was not, and it is more informative.
Attribute it. "A 2024 cohort study of 1,200 teenagers found…" places the claim's weight on a specific piece of work and lets the reader weigh it.
Own it as reasoning. If the claim is your inference from evidence you have already cited, write it that way. "Taken together, these findings suggest…" is a legitimate move that needs no citation, as long as the findings underneath are cited and the inference is reasonable.
Cut it. Often the fastest and best option. Read the paragraph without the sentence. If nothing is lost, it was padding wearing the costume of a fact.
What you should not do is keep searching until you find something that can be bent into shape. That is citation shopping, and it is exactly how quotation errors get into published papers by professional researchers. A citation that does not support its claim is worse than no citation at all: it looks like evidence, so nobody questions the sentence, and it collapses the moment one reader opens the paper.
Where the chatbot fits
A model is good at the parts of this that are about language. Ask it what researchers call your concept, which fields study it, who the well-known authors are, what search terms to try. That is real help, and it is the part where being wrong costs you nothing because the search results correct it immediately.
Do not ask it for the reference itself and paste the answer in. Asked to support a sentence, a model generates the citation that fits your sentence, which is a different task from finding the citation that exists. That is the mechanism behind fabricated references, and it is also why a reference that arrives already perfectly matched to your claim deserves more scepticism than one you had to dig for. Every reference from a chatbot still goes through the existence check and then the support check.
Next time, reverse the order
The whole procedure above exists because the sentence came first. When you have time, read first: gather sources, note what each one actually establishes, and write claims your notes already support. It is slower on day one and much faster at the end, because nothing needs retrofitting and nothing needs deleting the night before submission. Our guide to finding peer-reviewed articles covers the search side, and how to find and judge sources covers judging what you find.