TL;DR
- There is no universal AI citation-accuracy percentage because independent audits measure different units, engines, dates and failure definitions.[1]
- SourceCheckup found approximately 30% of individual GPT-4o Web Search medical statements unsupported, while nearly half of responses were not fully supported.[3][4]
- Oumi found that approximately 91% of assessed Gemini 3 AI Overviews contained the correct answer, but only 39% were both correct and fully supported.[5]
- A valid URL does not prove evidentiary support: Ahrefs found relatively low 404 rates, while separate studies still documented wrong attribution, unsupported claims and broader sourcing failures.[3]
- Citation trackers should separately measure URL resolution, publisher provenance, claim support and answer completeness instead of reporting one blended score.[1]
On this page
- What does “AI citation accuracy” mean?
- What did independent audits find for ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews?
- Why do AI citations fail?
- What did audits of Google AI Overviews and AI Mode actually measure?
- How should AI search visibility monitoring tools measure citation accuracy?
- What do the findings mean for users, publishers and AI visibility teams?
- Frequently asked questions
- References
AI citation accuracy has no single defensible percentage in 2026. Independent audits measure different outcomes—source identification, URL resolution, publisher attribution, claim support, or overall answer quality—so their figures must not be averaged into one failure rate.[1][2]
The clearest answer is that a visible citation is not proof of support. In one medical evaluation, approximately 30% of individual GPT-4o Web Search statements were unsupported, while nearly half of responses were not fully supported.[3][4] In Google AI Overviews, one audit found that approximately 91% of Gemini 3 overviews contained the correct answer, but only 39% were both correct and fully supported.[5]
This review compares what each audit actually tested, including its engine, sample, date, unit of analysis and headline result. It distinguishes a wrong article, a copied version, an unsupported claim, a fabricated URL and a legitimate redirect.
What does “AI citation accuracy” mean?
AI citation accuracy is not one metric. An audit may ask whether:
- The cited URL exists and resolves.
- The URL points to the article or publisher named by the engine.
- The cited page belongs to the original publisher.
- The page supports the specific statement attached to it.
- The answer accurately represents the source.
- Every claim in the answer is supported.
These tests produce different denominators. For example, Liu, Zhang and Liang reported average citation precision of 74.5% and citation recall of 51.5% in a 2023 evaluation of generative search engines.[6][7] Precision asked whether citations supported their associated statements; recall asked whether generated statements were fully supported by citations.[7][6]
A study reporting 11.0% of claims as inconsistent with cited evidence is therefore not directly comparable with a study reporting 31% of news responses with significant sourcing issues.[8] The first is claim-level; the second includes broader sourcing problems, including missing direct sources.[9]
What did independent audits find for ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews?
| Study | Publisher | Engines | Sample | What was checked | Headline finding as published | Date |
|---|---|---|---|---|---|---|
| How ChatGPT Search (Mis)represents Publisher Content | Columbia Journalism Review / Tow Center | ChatGPT Search | 200 quotes from 20 publishers | Publisher, URL and article-date identification | 153 responses were partially or entirely incorrect; the system acknowledged inability to answer accurately seven times.[10] | November 27, 2024[10] |
| AI Search Has a Citation Problem | Columbia Journalism Review / Tow Center | ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Copilot, Grok-2, Grok-3 beta and Gemini | 1,600 queries: 20 publishers × 10 articles × 8 engines | Original publisher, headline, date and URL identification | More than 60% of queries received incorrect answers.[2][11] | March 6, 2025[2] |
| Evaluating Verifiability in Generative Search Engines | Liu, Zhang and Liang / EMNLP | Bing Chat, NeevaAI, Perplexity and YouChat | Diverse search and open-ended questions; exact total not stated in the landing-page summary | Statement-level citation support | Average citation precision was 74.5%; average citation recall was 51.5%.[6] | December 2023[6] |
| An automated framework for assessing how well LLMs cite relevant medical references | Nature Communications / SourceCheckup | Seven LLM configurations, including GPT-4o Web Search and Gemini Ultra with retrieval | 800 questions and 58,000 statement-source pairs | URL validity and medical statement-source support | Between 50% and 90% of responses were not fully supported, and some were contradicted by cited sources.[4] | April 16, 2025[4] |
| Oumi AI Overviews study | Oumi | Google AI Overviews using Gemini 3 | 1,609 queries and 4,876 claims | Answer correctness and citation support | Approximately 91% contained the correct answer; only 39% were correct and fully supported.[5] | April 14, 2026[5] |
| Measuring Google AI Overviews | Xu, Iqbal and Montgomery | Google AI Overviews | 55,393 query executions; 7,583 overviews; 98,020 claims from 7,491 verifiable overviews | Claim consistency with cited evidence | 11.0% of claims were inconsistent with cited evidence; 41.9% of overviews were perfectly grounded.[12] | Data collected March 13–April 21, 2026[13][12] |
| Does the Citation Matter? | Agnoli et al. | Medical Google AI Overviews | 3,611 questions; 3,474 usable overviews; 12,175 scorable cited statements | Medical statement grounding | Observed statement support was 47.1%; the classifier-error-corrected estimate was 56.7%.[1] | October 29, 2026[14] |
| BBC news-integrity audit | BBC | ChatGPT, Copilot, Gemini and Perplexity | 100 questions and 362 reviewed responses | Accuracy, sourcing, impartiality and representation | 51% of all responses had significant issues; more than 45% of Gemini responses had significant sourcing errors.[9][15] | Testing in December 2024; report February 11, 2025[9] |
| International BBC/EBU study | BBC and European Broadcasting Union | Multiple assistants | 2,709 core-question responses and 353 custom-question responses, reviewed by 271 journalists in 22 organizations and 18 countries | News accuracy, sourcing and representation | 31% had significant sourcing issues; 45% had at least one significant issue of any kind.[15] | Responses generated May 24–June 10, 2025; report October 22, 2025[15] |
| How Often Do AI Assistants Hallucinate Links? | Ahrefs | ChatGPT, Perplexity, Copilot, Gemini, Claude and Mistral | 16 million unique URLs across two datasets | Whether cited or visited URLs returned HTTP 404 | ChatGPT returned 404s for 1.01% of visited URLs and 2.38% of cited URLs in the separate monitoring dataset.[16] | September 2, 2025[16] |
| Where do AI citation links lead? | Parse | ChatGPT, ChatGPT Search, Google AI Overviews and Google AI Mode | 248,227 addresses representing 8,254,220 citation events | Redirect chains and final destinations | Among 157,154 checks with a verdict, 16.7% did not land on the exact cited address.[17] | August 29, 2026[17] |
Why do AI citations fail?
1. Wrong article or destination
An engine may identify the correct publisher but link to the wrong article, a homepage or an unrelated page. The Tow Center’s eight-engine study evaluated whether systems could identify a source article from an excerpt, rather than merely whether a URL opened.[2]
The study reported that ChatGPT incorrectly identified 134 of 200 articles in its test, while 154 of 200 Grok 3 citations led to error pages.[2] These figures describe that controlled attribution task, not the error rate for all ordinary questions.
2. Syndicated or copied content instead of the original
A page can contain the relevant text and still credit the wrong publisher. In the 2024 ChatGPT Search study, ChatGPT cited a DMS Retail copy of New York Times material instead of the original publisher.[10]
The 2025 Tow Center study also found Perplexity Pro citing syndicated Texas Tribune articles in three of ten relevant queries, and ChatGPT citing a Yahoo News republication of a USA Today article.[2] This is a provenance failure, not necessarily a failure of topical relevance.
3. Unsupported attached claim
The most important failure is often subtler: the cited page exists and concerns the topic, but does not support the exact statement attached to it.
SourceCheckup’s GPT-4o Web Search results found approximately 30% of individual medical statements unsupported and nearly half of responses not fully supported.[3][4] In its HealthSearchQA subset, citation URL validity was 100%, but statement-level support was 75.7% and response-level support was 38.4%.[3]
Google AI Overviews show the same distinction. Oumi found that approximately 91% of Gemini 3 overviews contained the correct answer, yet only 39% were both correct and fully supported.[5] A correct answer can therefore contain additional claims that its citations do not establish.
4. Fabricated or unavailable URL
A fabricated URL is different from a real page that has moved, been deleted or blocked. Ahrefs explicitly warns that a 404 may represent a genuine page that is no longer available rather than an invented link.[1]
In the Tow Center’s Grok 3 test, more than half of Gemini and Grok 3 responses cited fabricated or broken URLs, and 154 of 200 Grok 3 citations led to error pages.[2] The audit combined fabricated and broken destinations in parts of its reporting, so those results should not be rewritten as a pure fabrication rate.
5. Redirected or noncanonical address
A citation may resolve to a different URL on the same site because of a redirect, URL normalization or a page move. Parse found that 83.3% of verdict-bearing checks landed on the exact cited address, 13.6% landed at a different address on the same website, 1.7% at a different website and 1.3% at a confirmed-gone destination.[17]
Its 16.7% figure applies to the 157,154 checks with a verdict, not all collected addresses.[17] A changed URL is not automatically an inaccurate citation: the final page may still contain the relevant evidence. The reverse also applies, because a page that resolves can still be unreadable to the engine: Finseo’s own crawler audit found that 8.9% of 20,293 pages across 524 sites returned nothing readable to AI crawlers, most of which never run JavaScript.
Why isn’t a cited link the same as a supporting source?
A cited link proves, at most, that the engine attached or exposed an address. It does not prove that the page contains the cited claim, that the claim is stated with the same qualifications, or that the page is the original publisher.
A proper support check compares the answer’s specific statement with the relevant passage on the final destination. It should also record whether the page is accessible, whether the publisher is the expected one and whether a redirect changes the source identity. This is why URL validity, publisher identity and claim entailment should be separate fields in an audit.[11]
What did audits of Google AI Overviews and AI Mode actually measure?
The 2026 Google studies are useful precisely because they report different units.
Oumi evaluated Google AI Overviews against SimpleQA questions and separated answer correctness from support. Its approximately 91% correct-answer figure does not mean that 91% of claims were supported.[12][18]
Xu, Iqbal and Montgomery split 98,020 claims from 7,491 verifiable overviews into categories including clear support, vague support, ambiguity, incorrect evidence and omission. They reported 11.0% of claims as inconsistent with cited evidence and 41.9% of overviews as perfectly grounded.[12] “Inconsistent” includes more than direct contradiction because the classification also considers omitted and ambiguous evidence.[12]
Agnoli et al. focused on medical questions and found 47.1% observed statement support, rising to an estimated 56.7% after correcting for classifier error.[1][14] Only 50.6% of all statements were scorable against citations, so this is not a support rate for every statement in every Google query.[1]
None of these studies establishes a claim-support rate for Google AI Mode. Parse included AI Mode in its destination-resolution audit, but that study tested where citation links led, not whether those destinations supported the answers.[17]
How should AI search visibility monitoring tools measure citation accuracy?
A citation tracker should retain the evidence chain rather than reduce it to a mention count.
- Save the complete answer and the exact claim associated with each citation.
- Store the displayed URL and any structured citation metadata.
- Resolve redirects and save the final destination.
- Check HTTP status separately from access blocks and bot challenges.
- Compare the final page with the expected publisher and article.
- Test whether the page supports the exact attached claim.
- Classify support as clear, partial, ambiguous, unsupported or contradicted.
- Record whether the source is original, syndicated, copied or unattributed.
- Preserve the engine, model, market, prompt, date and answer version.
- Report claim-level, citation-level and answer-level denominators separately.
OpenAI’s documentation distinguishes inline URL citations from a broader list of sources consulted by the system.[19] Perplexity’s API guidance similarly says to read URLs and metadata from the structured response payload rather than from URLs written in the model’s prose.[20] These distinctions matter when a tracker counts citations.
The result should be a matrix, not a single score:
| Dimension | Question |
|---|---|
| Existence | Does the URL resolve? |
| Destination | Where does the redirect chain end? |
| Publisher | Is the expected publisher credited? |
| Provenance | Is the page original or syndicated? |
| Relevance | Does the page address the topic? |
| Support | Does it substantiate the attached claim? |
| Completeness | Are all material claims cited? |
| Stability | Does the result remain valid over time? |
What do the findings mean for users, publishers and AI visibility teams?
For users, the first practical rule is to open the cited source. The second is to read the passage rather than assume that topical relevance equals support. A source about head injuries, for example, does not substantiate a specific treatment instruction unless the page actually provides that instruction.[1]
For publishers, being cited is not equivalent to being correctly attributed. A copied or syndicated page may receive the visibility while the original publisher loses credit.[2]
For SEO and AI-visibility teams, citation counts should be split into at least three measures: being mentioned, being cited and being supported. A brand may improve its citation count while still appearing beside claims that are incomplete, misattributed or weakly evidenced.
For audit designers, the main lesson is methodological restraint. The 2023 generative-search evaluation, the medical SourceCheckup study, the Google AI Overview audits, the BBC/EBU news studies and the URL-resolution audits are answering different questions.[3] None should be used to create a universal ranking of ChatGPT, Perplexity, Google, Gemini, Copilot or Grok without matching the task, date, sample and definition.
Frequently asked questions
How often does ChatGPT cite sources that do not support its answers?
There is no single rate for ChatGPT across all tasks. SourceCheckup found approximately 30% of individual GPT-4o Web Search medical statements unsupported, while nearly half of responses were not fully supported.[3][4] The Tow Center’s separate attribution test found 153 of 200 ChatGPT Search responses partially or entirely incorrect, but that test focused on identifying publisher, article and URL from excerpts rather than general claim entailment.[10][2]
Are Perplexity citations more accurate than ChatGPT citations?
The reviewed audits do not support a universal ranking. Liu, Zhang and Liang reported average citation precision of 74.5% across the systems they tested, including Perplexity, but that was a 2023 generative-search evaluation with its own task and definitions.[6] The Tow Center also found attribution and syndication errors in Perplexity and Perplexity Pro, while the BBC/EBU study measured news-sourcing issues rather than general citation support.[9][2]
Does a 404 prove that an AI fabricated a citation?
No. A 404 can result from a real page being deleted, moved or temporarily unavailable.[17] A reliable audit should distinguish confirmed fabrication, a dead page, an access failure, a redirect and a valid final destination.
What is the difference between citation precision and citation recall?
Citation precision asks whether the citations support the statements they accompany. Citation recall asks whether generated statements are fully supported by citations.[7] Liu, Zhang and Liang reported average precision of 74.5% and recall of 51.5% in their evaluation, showing why the two metrics should not be combined.[7][6]
Does Google AI Mode have the same citation accuracy as AI Overviews?
The reviewed audits do not establish that. Google AI Overviews were assessed for claim support by Oumi, Xu et al. and Agnoli et al., while Parse assessed AI Mode as part of a destination-resolution study.[11][17] A link reaching the expected destination does not show that the page supports the attached claim.
How can a citation tracker check whether an AI citation is accurate?
It should capture the answer, attached claim, displayed URL, redirect chain, final destination, publisher identity and support verdict. It should report separate rates for URL resolution, original-publisher attribution, claim support and answer-level completeness rather than one blended accuracy score.[1]
Do the reviewed audits cover brand visibility in Claude answers?
Among the reviewed audits, Claude appears in Ahrefs' study, which checked whether cited or visited URLs returned HTTP 404 across 16 million unique URLs.[16] That study measured link availability, not whether the cited pages supported the attached claims, and Ahrefs warns that a 404 may reflect a genuine page that is no longer available.[1]
Which engines did Parse test for citation link destinations?
Parse tested ChatGPT, ChatGPT Search, Google AI Overviews and Google AI Mode, covering 248,227 addresses that represented 8,254,220 citation events.[17] Among 157,154 checks with a verdict, 83.3% landed on the exact cited address and 16.7% did not.[17] The study tested where links led, not whether those destinations supported the answers.[17][2]
Should AI visibility teams track mentions, citations or supported citations?
Citation counts should be split into at least three measures: being mentioned, being cited and being supported. Being cited is not equivalent to being correctly attributed, because a copied or syndicated page may receive the visibility while the original publisher loses credit.[2] Trackers should report URL resolution, publisher provenance, claim support and answer completeness separately.[1]
References
- 1Does the Citation Matter? Quality, Grounding, and Physician-Rubric Accuracy of Google’s AI Overviews for Medical Information.arxiv.org
- 2AI Search Has a Citation Problem - Columbia Journalism Review.cjr.org
- 3An automated framework for assessing how well LLMs cite relevant medical references | Nature Communications.nature.com
- 4An automated framework for assessing how well LLMs cite relevant medical references - PubMed.pubmed.ncbi.nlm.nih.gov
- 5Oumi’s Study Finds 50% of AI Overviews Untrustworthy – Oumi Blog.blog.oumi.ai
- 6Evaluating Verifiability in Generative Search Engines - ACL Anthology.aclanthology.org
- 7citations. However, not all generated statements are.aclanthology.org
- 8News Integrity in AI Assistants.bbc.co.uk
- 9Ebu Mis Bbc News Integrity In Ai Assistants Report 2025.ebu.ch
- 10How ChatGPT Search (Mis)represents Publisher Content - Columbia Journalism Review.cjr.org
- 11Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents.arxiv.org
- 12Measuring Google AI Overviews:Activation, Source Quality, Claim Fidelity, and Publisher Impact.arxiv.org
- 13WashU research finds gaps in Google AI Overview citations, claims - The Source - WashU.source.washu.edu
- 14[2610.39339] Does the Citation Matter? Quality, Grounding, and Physician-Rubric Accuracy of Google's AI Overviews for Medical Information.arxiv.org
- 15News Integrity in AI Assistants.bbc.co.uk
- 16New Study: How Often Do AI Assistants Hallucinate Links? (16 Million URLs Studied).ahrefs.com
- 17Where do AI citation links lead? | Parse.parse.gl
- 18Representation of BBC News content in AI Assistants.bbc.co.uk
- 19Web search.developers.openai.com
- 20Prompt Guide - Perplexity.docs.perplexity.ai