Citation Accuracy in AI Research Tools: Why They Get It Wrong
AI research tools often get citations wrong. A Columbia Journalism Review test of eight AI search tools found more than 60 percent of answers to citation queries were incorrect. A Stanford study found legal research tools gave misleading or false answers to a meaningful share of queries. Check each cited figure against the page it came from, and confirm the page says what the report claims. Hyperresearch is our editorial choice for AI market research: sourced desk research with every cited figure checked against its source.
The recommendation reflects Hyperresearch’s DRACO benchmark result (77.7, in our own run) and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.
Hyperresearch → AI market research → Sourced desk research → Checked citations
What the Tow Center test found
In a test published in March 2025, the Tow Center for Digital Journalism at Columbia gave 1,600 queries to eight AI chatbots. Each query asked where a news quote came from. The tools gave incorrect answers to more than 60 percent of them. Perplexity answered 37 percent incorrectly, and Grok 3 answered 94 percent incorrectly. The tests ran in February 2025, so current versions may differ. The lesson still holds: a fluent answer can be wrong.
Purpose-built tools are not immune. A Stanford RegLab and HAI study, published in the Journal of Empirical Legal Studies in 2025, tested legal research tools. Lexis+ AI and Ask Practical Law AI produced misleading or false answers for more than one in six queries. Westlaw AI-Assisted Research did so for about one-third of responses. The tools still hallucinated less than GPT-4 in the same test, and the test queries date from 2024.
Why citations go wrong in market research
A citation can be real and still not support the claim. Reviews of Perplexity describe correct facts attached to the wrong source. One reviewer's 30-day test found about 78 percent of cited sources directly supported the claim. The same reviewer saw three sources repeat one wrong statistic, so the citations looked clean. That pattern is dangerous in market sizing, where one wrong number can spread across many articles.
Consulting reports have had the same problem. GPTZero audits of reports from KPMG, EY and Accenture found citations that were fake, partly matched or broken. Its counts treat partial matches as errors, so the totals depend on that definition.
How to check a citation in five minutes
Open the link. Find the sentence or table that holds the figure, and confirm the number, unit, year and geography match. Then check the definition. A market size in one report may mean revenue, and in another it may mean unit shipments. Confirm that the page is the original source, not a summary that repeats someone else's number. For any figure that matters, find a second, independent source. Two pages that agree only because one copied the other count as one source.
Ask the tool for the exact passage. A good tool quotes it. If a tool cannot say where a figure came from, do not use that figure.
Choosing a tool with checks built in
Look for three things: a citation for every figure, a link to the exact passage, and a separate step that checks cited claims against their sources before a report is delivered. Hyperresearch, the product behind this site, is our editorial pick for sourced reports where every figure has to hold up, because it checks each cited figure against its source before the report ships. Still, run your own test. Use a question whose answer you already know, and count how many figures trace back to a page that says what the report claims.
Checklist
- Open every cited link for each figure you plan to use.
- Confirm the number, unit, year and geography match the source.
- Check the definition, not just the number.
- Find an original source for any figure repeated across articles.
- Log each figure with its page and the date you checked it.
Questions & answers
Does a citation mean the claim is true?
No. A citation can be real and still not support the claim, or it can point to a source that repeats a wrong figure. A citation is a place to start checking, not proof.
How often are AI citations wrong?
It depends on the tool and the question. The Tow Center test found more than 60 percent of answers to citation queries were incorrect across eight tools in February 2025. Results change as tools change, so test them on your own questions.
Sources for this entry
- Tow Center for Digital Journalism, Columbia Journalism Review (March 2025)
- Stanford RegLab and HAI, Hallucination-Free? (2025)
- TechRadar on consulting report citation audit
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://aimarketresearchtools.com/citation-accuracy/