Search the exact title in quotation marks, look for a working link or a DOI, and confirm the publication venue is real before you accept anything the citation is attached to. If any one of those checks fails, treat the citation as unverified rather than assuming a typo or a formatting slip on the AI's part. AI tools invent citations by generating text in the general shape of a real one, which means the check has to confirm existence first, and accuracy of the claim second. Skipping straight to "does this claim sound right" without confirming the source exists is the single most common way a fabricated citation slips into finished work.
Step 1: Search the exact title, in quotes
Copy the study or report title the AI gave you exactly as written, wrap it in quotation marks, and search it. This forces an exact-match search rather than a loose one, which matters because a search engine's "did you mean" suggestions can quietly point you toward a real, similarly named source and give a false sense that the citation checked out.
- Nothing comes up at all: strong warning sign. Treat the citation as unverified.
- A close but not exact match comes up: the AI may have blended two real sources, or altered a real title slightly. Investigate further before accepting it.
- An exact match comes up: move to the next step. Existing under the right title is necessary, but not sufficient.
Step 2: Follow the link, DOI, or database entry directly
A real citation should lead somewhere. If a DOI was provided, paste it into doi.org directly and confirm it resolves to a page with a matching title and author. If a direct link was given, open it and confirm the content matches, rather than assuming the domain alone makes it legitimate. If neither exists, search the title in a database appropriate to the field, such as Google Scholar for academic work, or the issuing organization's own publications page for an industry report.
A citation with no way to independently locate the actual document is not something you can verify. It stays unverified, no matter how plausible the surrounding details sound.
Step 3: Check that the publication details actually line up
Once you have located what looks like the right source, check the specifics rather than trusting the general shape:
- Does the publication venue named actually exist, and does it publish this kind of work?
- Do the volume, issue number, or publication date align with that venue's actual history?
- Is the named author actually associated with this piece, not just a plausible-sounding name in the field?
A mismatch on any of these, even a small one, is worth treating as disqualifying until resolved, not waved through as a minor detail.
Step 4: Confirm the source actually says what the AI claims
This is the step people skip once they have confirmed a source exists, and it is where a second, sneakier kind of hallucination lives: a real study, cited correctly, attached to a claim it never actually makes. Open the source and read the specific section relevant to the claim. Do not rely on an abstract or summary alone if the claim is central to your piece; abstracts can omit caveats, sample limitations, or qualifications that change what the finding actually supports.
What a passing check looks like, end to end
- The title exists under an exact search.
- A DOI, direct link, or database entry resolves to the actual document.
- The publication details, author, venue, date, line up with that source's real history.
- The source, read directly, actually supports the specific claim attached to it.
A citation that passes all four is verified. A citation that fails any single one is unverified, and unverified means it gets cut or clearly flagged, not softened into a hedge and left in place. See how to verify a statistic before you publish it for the companion check once a source's existence is confirmed but its numbers still need scrutiny.
How much time this actually takes
For a single citation, the four-step check usually takes a few minutes: one search for the exact title, one click to follow the link or DOI, a quick scan of the publication details, and a read of the relevant section. The time adds up on a heavily cited draft, which is exactly why step 2 in how to fact-check anything an AI writes for you exists: rank claims by how much a wrong one would cost before you start verifying, so the time goes to the citations that actually matter to your piece's argument, not to every passing mention equally.
A shortcut that does not work
Asking the AI tool "is this citation real?" is not a substitute for the four steps above. The same process that produced the original citation produces the answer to that follow-up question, with no new verification step added in between. Treat a self-reported confirmation from the tool as no more reliable than the original claim, because structurally, it is not.
FAQ
What if the citation exists but I cannot access the full text?
Look for the abstract, a summary from the publisher, or a citing article that quotes the relevant section directly. If you genuinely cannot confirm what the source says, the honest move is to say so in your own piece, or to leave the claim out, rather than repeating the AI's characterization of a source you were unable to check.
Does a citation from a well-known publication need the same scrutiny?
Yes. A well-known publication name does not confirm that the AI accurately represented what that publication said, or even that the specific article exists at all. The reputation of the venue does not transfer to an unverified claim about what it published.
How do I check a citation for a general web article rather than an academic study?
The same four-step logic applies: confirm the article exists under that title, at that outlet, follow the actual link, and read what it says rather than trusting a paraphrase. Web articles are generally easier to verify than academic papers because there is usually a direct, public link to check.
Is it faster to just remove any claim I cannot quickly verify?
For lower-stakes claims, often yes, and that is a reasonable editorial choice. For claims central to your piece's argument, cutting rather than verifying trades a fact-checking problem for a credibility problem, since the piece may end up making a weaker case than the evidence would actually support.
Can AI tools with web search built in still invent citations?
Yes, though generally at a lower rate than a model working from memory alone. Built-in search reduces the odds of a fully invented source, but it does not guarantee accurate representation of what a real source says, so the same verification steps still apply.
Should I keep a record of which citations I checked and how?
For any piece with more than a few citations, yes. A short internal note of what you searched, what you found, and whether it passed all four steps saves time if a claim is later questioned, and it makes the check itself more disciplined than trying to hold the whole process in your head at once.