Broken links in AI-generated citations: how to check them
A citation whose link does not work is not a citation you can defend. Learn to tell a dead link from a moved page or a fabricated reference, and what to do with each.

In short: a broken link is a failed citation, not a minor flaw. Click every link, title-search the dead ones, and keep only the citations you could actually reach and match to your claim.
Why broken links matter more than they look
A broken link hides three very different problems, and each needs a different fix:
- A real paper whose link moved. Journals migrate platforms, preprint servers renumber pages, and DOIs get re-registered. The paper exists; only your route to it is stale.
- A real paper cited through a wrong link. A typo, a copied DOI from a different paper, or a URL pasted from a search snippet can send a reader to the wrong landing page. The source exists, but the link does not prove it.
- A reference that was never real. A complete-looking citation — authors, journal, year, DOI-style string — that resolves to nothing, anywhere. Until proven otherwise, treat it as fabricated, not merely mislinked.
You cannot tell these apart by looking at the reference. You tell them apart by checking, in order.
What the pilot found (pilot numbers, not a universal rate)
In a published pilot audit of 66 AI-generated citations (Paper 2, study 1 — a pilot, not a universal rate and not a vendor test), 31 citations (47%) had broken links. A further 7 (11%) mismatched the claim they were attached to, and 11 (17%) could not be verified. Only 5 citations both resolved and fully supported the claim they were attached to.
The point of those numbers is not that every AI draft fails at the same rate. It is that a broken link is the single most common failure mode in AI-generated references, and also the easiest to catch — if you click every link instead of sampling a few.
The check, link by link
1. Click every link, not a sample. Open the DOI or URL for each AI-assisted citation. A dead link, a “page not found”, or a DOI that resolves nowhere is a failed citation at this step, even if the rest of the reference looks perfect. Record it as Broken and move to step 3.
2. Check that the landing page matches the reference. A link that resolves is not automatically a pass. Compare the title, authors, year, and journal on the page you reached against the reference in your draft. If the landing page is a different paper, a journal homepage, or a generic search page, the link does not support your citation — treat it as a wrong-link failure, not a pass.
3. If the link is dead, search by title before you give up. Paste the exact title into Crossref, OpenAlex, Semantic Scholar, or Google Scholar, and check the publisher's own site. If you find the real paper with a working DOI or URL, replace the dead link with the correct one and re-run steps 1–2. If no record of the paper exists in any of these places, you have not found a moved page — you have found a reference you cannot stand behind.
4. Separate “blocked” from “broken”. A page behind a paywall, a login wall, or a temporarily unreachable server is not the same as a dead link. Mark it Unverifiable, not Broken: you could not confirm it, so you cannot cite it on trust either. Try your library catalogue or the publisher's site as a second route; if it still cannot be reached, it still cannot stay in the paper as written.
5. Keep a one-line record per citation. Citation, link result (resolves / dead / wrong landing page / blocked), and the source you actually opened. This takes under a minute per citation and is your answer if a reviewer asks how you checked.
Free tools that help (and what they do not do)
- Crossref DOI resolver: tells you whether a DOI exists and where it points. A DOI that does not exist in Crossref is a strong sign the reference was generated, not retrieved.
- OpenAlex and Semantic Scholar: open bibliographic records — use them to confirm a paper's real title, authors, year, and DOI.
- Google Scholar and the publisher's site: the final word when an index and your draft disagree. The publisher's page outranks every index.
- What no tool does for you: judge whether the source supports your sentence. A link that resolves to a real paper can still be a mismatch — the paper may say something entirely different from your claim. Resolving the link is step one; reading the relevant passage is step two.
What to do with each result
- Dead link, real paper found elsewhere: replace the link with the correct DOI or URL, and record the correction.
- Dead link, no record found anywhere: drop the citation. If the claim matters, find a real source for it or drop the claim too. Never keep a citation you could not verify.
- Wrong landing page: fix the link if the intended paper is identifiable; otherwise treat it like a dead link and drop it.
- Blocked or unverifiable: do not cite it on trust. Find an accessible copy through your library, or replace it with a source you could actually open and read.
Common traps
- The confident fake. A complete-looking reference that resolves to nothing. Always click — appearance is not evidence.
- The near miss. A link that resolves to a real paper, but the wrong one. Always compare title and authors on the landing page.
- The homepage redirect. A dead article URL that quietly redirects to a journal's homepage looks like a working link. It is not: it proves nothing about your citation.
- The borrowed DOI. A real DOI attached to the wrong paper, copied from another reference list. The DOI resolves, the citation still fails.
- “It worked when the draft was written.” Links rot after drafting. Re-run step 1 on every link on the day you submit, not just the day you drafted.
A 10-minute routine for a batch of ~10 citations
Minutes 1–4: click every link; mark dead links and wrong landing pages. Minutes 5–7: title-search the dead ones; replace the links you can genuinely recover. Minutes 8–9: check blocked ones once more via your library or the publisher. Minute 10: write the one-line record. A paper with 60 AI-assisted citations takes about an hour — less time than answering one reviewer query about a reference that leads nowhere.
Questions, answered
Is a broken link a minor flaw?
No. A broken link is a failed citation: if a reader cannot reach the source, they cannot check your claim, and a reviewer will treat the reference as unproven.
How do I tell a dead link from a moved page or a fabricated reference?
Click the link, then title-search any dead link in Crossref, OpenAlex, Semantic Scholar, or the publisher's site. If you find the real paper with a working link, replace the dead link; if no record exists anywhere, treat the reference as one you cannot stand behind and drop it.
How common are broken links in AI-generated citations?
In a published pilot audit of 66 AI-generated citations, 31 citations (47%) had broken links (Paper 2 v2, Trust, But Verify, DOI 10.5281/zenodo.23731235). That is a pilot result, not a universal rate.
Sources and further reading
- Paper 2 v2, Trust, But Verify: A Claim-Level Citation Audit of AI-Generated Strategic Research Briefs — Zenodo DOI 10.5281/zenodo.23731235 (pilot audit of 66 citations, study 1).
- Working Paper v3, From Answers to Auditable Intelligence — Zenodo DOI 10.5281/zenodo.23718472 (the framework behind the check; a design proposal, not validated by a blinded pilot).
- How to check AI-generated citations in your own paper — the full four-verdict method this guide draws on.
Oryx Centre for Strategic Futures is a two-person research initiative (Udyam-registered micro enterprise) working on auditable AI. Its Oryx AI pilot reports are payable only on acceptance, currently invite-only beta.