Are ChatGPT citations accurate? What the one measured study found, and what to do when you are cited wrong

On 6 March 2025 the Tow Center for Digital Journalism published the one study we trust on this question.

Klaudia Jaźwińska and Aisvarya Chandrasekar ran 1,600 queries through eight AI search tools, and CJR reported that the tools gave incorrect answers to more than three of every five of them (CJR, AI Search Has a Citation Problem, read on 14 October 2026).

That number gets quoted everywhere.

It also gets quoted wrong, and that is the complication. The study did not measure whether the answers were true. It measured whether the tools could name the right source.

For a business, that second failure is the one that hurts.

Step 1: What the study tested

The design was narrow on purpose.

The researchers took 20 news publishers, picked ten articles from each, and pasted a direct excerpt into each of eight tools: ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, Grok-2, Grok-3 and Gemini, per the CJR write-up of 6 March 2025.

Then they asked each tool for the headline, the publisher, the publication date and the URL.

That is a fair test of citation. The excerpt is verbatim, so a tool that searched properly should find the original. Getting the source wrong here means the attribution layer failed, not the writing.

Step 2: The numbers, with their dates

Every figure below comes from that one CJR article, published 6 March 2025 and read on 14 October 2026.

In the Tow Center study, ChatGPT Search misidentified 134 of its 200 articles, signalled uncertainty only 15 times, and in the authors' words "never declined to provide an answer."

Perplexity got 37 of every 100 queries wrong, according to the same study.

In the same CJR study, Grok 3 got 94 of every 100 wrong, and the piece reports that 154 of its 200 citations led to error pages.

More than half of responses from Gemini and Grok 3 cited fabricated or broken URLs, the study found.

And the paid tiers did worse in one way that matters. The authors write that premium models answered more prompts correctly but "paradoxically also demonstrated higher error rates," because they gave confident wrong answers instead of declining.

So the short answer to "are ChatGPT citations accurate" is no, not reliably.

But read the limits before you repeat any of it.

Step 3: What the study cannot tell you

The authors say themselves that their "research design may not reflect typical user behavior" and that the findings "are not intended to be extrapolated to all models or news organizations."

Three more limits we would add.

It is old. The tests ran in early 2025. ChatGPT now runs on a different model family, and our GPT-6 citations post covers how much changed in October 2026 alone.

It used news. A news article has one clear original. A local business page is often paraphrased by many directories, which makes the "right source" harder to define, not easier.

It measured one task. Pasting an excerpt and asking for the source is a lab test. Real users ask open questions, and the citation sits beside a generated sentence that may or may not match it.

So nobody can tell you today's error rate. We cannot either.

What we can say is that the failure was measured, it was large, and the vendor has since put a warning in writing.

Step 4: What OpenAI says about its own citations

OpenAI's help page on searching the web with ChatGPT, which we opened in a browser on 14 October 2026, is blunt.

It says "Search results and citations can be incomplete, outdated, or incorrect."

Then it tells users to open a cited source to check that it supports the answer, review when it was published or updated, and use an authoritative source when accuracy matters.

That is the vendor telling its users to verify the thing your business depends on.

The same page says responses that use web search "may include citations," and that you can select Sources "when available." It also says placement in results is not guaranteed, and that a site has to allow OAI-SearchBot to be eligible at all.

Nothing on that page promises the citation supports the sentence beside it.

Step 5: The two ways this lands on a business

The CJR study was about publishers. The mechanics transfer.

Failure one: you are cited for something you never said. The answer states a price, a policy or a claim, and your URL sits next to it. A reader who never clicks assumes you said it.

Failure two: you are cited to the wrong page. The tool credits a directory, a scraper or a syndicated copy instead of your page. The Tow Center saw this with news: Perplexity Pro cited syndicated versions of Tribune articles for three of ten queries, per the 6 March 2025 piece.

Both failures are quiet.

You will not get an alert. You only see them if you run the prompts yourself.

Here is the unpopular part. Most of the fix is on your side.

The model is not reading your mind. It is reading whatever it retrieved, and the recommendation pipeline shows where a wrong page enters: retrieval picks a stale copy, extraction pulls the wrong attribute, composition writes a sentence the source never made.

Step 6: What to do when you are cited wrong

These are checks, not tricks. They work whether or not the error rate has improved after the 2025 tests.

1. Save the evidence. Copy the prompt, the full answer and the cited URL, with the date. A screenshot without the prompt is useless later.
2. Open the cited page. If it is yours, find the passage the model leaned on. Ambiguous wording, an old price left on an archive page, or a figure in an image are the usual suspects.
3. Fix your own page first. State the correct claim in plain text near the top. Put a visible updated date on it, the way VALORAE corrects a page in place and changes its date when a figure moves.
4. Chase the third-party copy. If a directory or a scraper is cited instead of you, ask that site to correct or remove it. If it is a copy of your content, make sure your original is crawlable and clearly the canonical version.
5. Rerun the prompt several times over the following weeks. One clean run proves nothing. The CJR authors note that responses "can vary in response to the same query," which is also why AI visibility tools disagree.

Add a description-accuracy column to whatever sheet you already keep. Our measurement guide has the full prompt set.

Step 7: Where NovaTechRay draws the line

We do not know how to make a model cite you correctly every time. Nobody does.

Anyone selling a guaranteed citation fix is selling something the vendor itself will not promise. OpenAI's help page says placement is not guaranteed and citations can be wrong.

What NovaTechRay does is narrower. We find the pages the assistants cite for your category, check whether they say what you want said, and fix the ones you control.

That is slower than a promise. It is also the only part anyone can verify.

The short version

• Treat ChatGPT citations as leads to check, not as proof.
• Quote the Tow Center study with its date, March 2025, and its scope, news excerpts.
• Run your money questions by hand and record which URL gets cited.
• Fix your own page first when an answer misquotes you.
• Ask third-party sites to correct copies that get cited instead of you.
• Rerun the same prompts for weeks before calling anything fixed.

This is the work NovaTechRay does at novatechray.com.