LLMs cite wrong sources even when answers are right
A new Peking University benchmark shows LLMs cite passages that do not support their answers, even when the answers are correct. Your citation dashboard is noisier than you think.
Key takeaways
- GPT and Gemini routinely cite passages that do not support their (correct) answers, per Peking University's CiteVQA benchmark.
- Brand citation share in LLMs likely overstates and understates actual influence at the same time, in unpredictable ways.
- Regulated sectors (finance, law, medicine) inherit the largest exposure because the citation is the work product.
- AEO tools that measure domain citations without verifying passage accuracy are answering half the question.
- Make institutional name and load-bearing statistics inseparable at the passage level, not just the document level.
What happened
Per The Decoder, researchers at Peking University have documented a failure mode they call "attribution hallucination": leading models including GPT and Gemini routinely answer document questions correctly, then point to passages that do not actually support the answer. The team built a benchmark, CiteVQA, to test the gap systematically. It is the first benchmark designed to separate answer accuracy from citation accuracy.
The headline finding: getting the right answer tells you almost nothing about whether the cited evidence is real. Models confidently fabricate the provenance of correct conclusions. The researchers flag the obvious danger zones, law and medicine, where the citation is the work product, not the answer.
For brands that depend on being cited inside LLM outputs, the implication runs the other way too. If models routinely misattribute correct answers to the wrong passages, then "we got cited by ChatGPT" is a weaker signal than the industry has been treating it as.
Why it matters for your brand
The B2B authority playbook of the last 18 months assumes a clean chain: model reads your research, model uses your finding, model cites your URL. CiteVQA breaks that assumption. The model may use your finding and cite a competitor. It may cite you for a claim you never made. It may cite the right document but the wrong page. The answer the user sees can be correct and your brand can still be misrepresented, or absent, in the citation slot.
For financial services, this is a compliance problem dressed as a marketing problem. If an LLM correctly summarises a Basel III provision but attributes it to a bank's marketing blog instead of the BIS, the bank inherits an authority it did not earn and a liability it did not price. The reverse is worse: a regulator's actual guidance summarised correctly but cited to a third party means the regulator loses the visibility it needs to function. Communications teams at central banks, the IMF, and large supervisors should assume their citation share underrepresents their actual influence on model outputs, and budget accordingly for direct distribution.
For multilaterals and UN agencies, attribution hallucination compounds an existing problem. These institutions produce the underlying data (UNDRR on disaster loss, WHO on mortality, World Bank on poverty lines) that gets restated by think tanks, news outlets, and consultancies. Models already tend to cite the restatement over the primary source. If the citation layer is also unreliable at the passage level, the primary source can do everything right (open licensing, structured data, clean PDFs) and still be edited out of the answer it generated. The fix is not better SEO. It is making the institution's name and the specific dataset name inseparable in the model's training and retrieval context.
For major industrial groups, the risk is sharper in technical content. A cement producer publishing decarbonisation methodology, an engineering firm publishing standards interpretation, a pharma company publishing trial results: in each case the value of the citation is the precision of the passage. "Holcim says" is worth less than "Holcim's 2025 sustainability report, page 47, says." If models cannot reliably hit the passage, the procurement officer or analyst using the LLM will either stop trusting citations or click through to verify. Either behaviour collapses the assumed funnel from AI answer to brand consideration.