Ask Perplexity a question and the answer arrives with numbered citations, each a link that opens. Perplexity searches before it writes, so the pages are usually real. Are Perplexity's citations accurate, then? The link usually is; the claim in front of it is the part to check.
A fake reference to a paper that does not exist is the ChatGPT failure. Perplexity's failure is a real page under a sentence it does not support. So checking Perplexity has two halves: reading each page, which only you can do, and confirming each reference against the publisher's record, which EdCitation's free Verify references does for a whole list. Two groups of publishers have also alleged in court that Perplexity attributed text to them that they never published.
How does Perplexity find and cite its sources?
Perplexity runs a web search for every question, then has a language model write an answer from the pages it retrieved, with a numbered citation after the sentences taken from each page. Its help centre says answers draw on "authoritative sources like articles, websites, and journals" and that each answer "includes numbered citations linking to the original sources" (checked on 22 September 2026). On the free plan a mode called Best picks the writing model; on paid plans you choose it.
Modes, models and plans
Perplexity's own comparison table puts standard search at one or two sources with "basic citations", and Pro Search at dozens of sources with direct links (checked on 22 September 2026). Deep Research "performs dozens of searches automatically, reads hundreds of sources" and writes a report in two to four minutes. The models page lists ten Search models on Pro and Max, from Perplexity, OpenAI, Google, Anthropic, xAI and others. Free accounts get three Pro Searches a day and one Research query a month; the Pro page showed Pro at $17 a month when billed annually.
What Perplexity itself says about accuracy
Perplexity does not claim its answers are always right. Its Pro Search page says it is "crucial to validate information by referencing the sources linked in your answer". Some citations now carry a Government, Academic or Trusted label; the label rates the whole domain, not the claim, and the help centre says a label "is never a substitute for reading the source yourself".
What have studies found about Perplexity's citations?
Three independent tests have measured Perplexity's citations directly; in all of them the links open and a meaningful share do not support their sentence.
| Test | When | What was measured | Perplexity result |
|---|---|---|---|
| Liu, Zhang and Liang (2023), Stanford | Feb to Mar 2023, 1,450 queries per engine | Citations supporting their sentence; sentences fully supported | Precision 72.7%, recall 68.7% (four-engine averages 74.5% and 51.5%) |
| Min et al. (2023), FActScore | 2023, biographies checked against Wikipedia | Share of facts supported | 71.5%; citations on 36.0% of supported and 37.6% of unsupported sentences |
| Jaźwińska and Chandrasekar (2025), Tow Center | Feb 2025, 200 news excerpts per chatbot | Naming the article, publisher, date and URL | Wrong on 37%; Perplexity Pro answered more but erred more |
| Özbek and Bağcıer (2026), Indian Journal of Orthopaedics | 2026, 3,150 references from three tools | Existence, bibliographic accuracy, PubMed ID, relevance | Worst hallucination score of the three, 6.51, against 1.81 for ChatGPT |
The Stanford audit: real links, unsupported sentences
Liu, Zhang and Liang (2023) had 34 trained annotators check 1,450 responses from each of four citing search engines, including perplexity.ai, collected in February and March 2023. Citation precision, the share of citations that support their sentence, was 72.7% for perplexity.ai, second to Bing Chat's 89.5%. Citation recall, the share of sentences fully supported by their citations, was the highest of the four at 68.7%, against an average of 51.5%. So 27% of its citations did not support their sentence, and 31% of its sentences were not fully supported.
Min et al. (2023) found that 71.5% of the facts in PerplexityAI's biographies were supported by Wikipedia, the best of three systems, but that its citations had "little correlation with factual precision": 36.0% of supported sentences carried a citation, and so did 37.6% of unsupported ones.
The Tow Center test: the paid tier answered more and erred more
The Tow Center for Digital Journalism gave eight chatbots excerpts from 200 news articles and asked each to identify the article, publisher, date and URL (Jaźwińska & Chandrasekar, 2025). More than 60% of answers were wrong. Perplexity answered 37% incorrectly, the lowest figure the report gives; Grok 3 94%. Perplexity Pro, then $20 a month, answered more queries correctly than the free version but had a higher error rate, because it declined less and gave "definitive, but wrong, answers". The free version also identified all ten excerpts from National Geographic, which had blocked its crawler, although Perplexity says it respects robots.txt; and Perplexity Pro cited syndicated copies of Texas Tribune articles, a partner, for three of ten queries.
The one peer-reviewed test of Perplexity's reference lists
The evidence from academic and medical settings is thin. Özbek and Bağcıer (2026) asked ChatGPT, Gemini and Perplexity for references on 30 rotator cuff subtopics in two manuscript formats, 3,150 references in all. Perplexity had the highest mean hallucination score, 6.51, against 4.01 for Gemini and 1.81 for ChatGPT, where higher is worse. The abstract does not name the model versions or search mode, and the paper is paywalled, so treat it as one result rather than a rate. A health librarian's review of the free modes in May 2025 found at least eight citations per answer, accuracy not guaranteed, and no visibility into how sources are chosen (Roy, 2025).
The largest 2026 study of citation URLs, Rao, Wong and Callison-Burch (2026), did not test Perplexity; in the ten search-backed systems it did test, 3% to 13% of cited URLs had probably never existed. The general studies of invented references, such as Walters and Wilder (2023), are tabulated in why AI tools invent references.
Why do Perplexity's citations fail when the links work?
Perplexity's citations fail because the search finds a page, the model writes a sentence, and nothing checks that the second is contained in the first.
The page does not say what the sentence says
This is what the Stanford audit measured. The commonest form is partial support: the page says something adjacent, the sentence says more, and the number in brackets makes the whole sentence look sourced. A related form is the synthesised sentence, written from several pages with one citation; Liu, Zhang and Liang (2023) note that the scope of a single citation marker after two statements is often ambiguous.
The summary cited instead of the paper
Perplexity cites the page its search ranked, and for a research finding that is often a news report, a university press release or a syndicated copy rather than the paper. The Tow Center saw this with news: syndicated versions on Yahoo News and AOL cited instead of the original. A press release gives the finding without its caveats, and a reference to it is not a reference to the study.
The finding stated more strongly than the paper states it
The Stanford audit found that engines which copied text closely from their sources scored higher precision, and perplexity.ai copied least. A fluent paraphrase reads better and drifts further from the page; "was associated with" becomes "causes". The invented quotation is the extreme case: in the Dow Jones complaint below, Perplexity is alleged to have produced "several key quotes" from a Wall Street Journal article of which only one phrase appeared in it. A request for "peer-reviewed references" in ordinary Search can also return journal articles mixed with blog posts and news, each with a working link.
Each failure has its own check, and a lookup catches only some of them:
| Failure | What catches it | Where EdCitation helps |
|---|---|---|
| The page does not say what the sentence says | Reading the passage | It cannot: no lookup reads for you |
| A summary cited instead of the paper | Finding the paper the summary reports | Find sources, searched by the finding; then Cite a source from its DOI |
| The finding stated more strongly than the paper states it | The paper's own wording | Find sources puts the paper in front of you to read |
| An entry in a reference list that does not exist | Looking every entry up | Verify references |
Has Perplexity attributed invented text to publishers?
Two groups of publishers have alleged in court that it has. These are allegations, not findings; they are cited because the complaints show the outputs.
Dow Jones and NYP Holdings, in a second amended complaint filed on 28 January 2025 in the Southern District of New York, say a Perplexity Pro user asked for the full text of a New York Post article and received its first 139 words verbatim followed by "five paragraphs of made-up text" attributed to the Post, and that a request for key quotes from a paywalled Wall Street Journal article returned quotes that do not appear in it (Dow Jones & Co. v. Perplexity AI, 2025). Encyclopaedia Britannica and Merriam-Webster filed a similar complaint on 10 September 2025, alleging that Perplexity "sometimes generates hallucinations in its outputs and attributes that text" to them (Encyclopaedia Britannica v. Perplexity AI, 2025).
For a student the lesson is narrow. A quotation attributed to a source by Perplexity is checked like any other claim: open the source and find the words.
How do I ask Perplexity for references I can check?
Ask for the identifiers and the exact passage, and tell it what to do when it cannot find either. These prompts do not make the answer accurate; they make it checkable.
- Ask for the paper, not the page. "For each source, give the DOI and the link to the paper on the publisher's site, not a news article, press release or summary." Where it offers only the press release, search for the paper yourself in EdCitation's Find sources.
- Ask for the supporting words. "For each sentence, quote the exact passage from the source that supports it, in the source's own words."
- Ask it to admit a gap. "If you cannot find a source that says this, write 'no source found' instead of a general statement."
- Use Research mode for scholarly work, select the Academic focus where it is offered, and check the sources it read in the progress panel.
No prompt stops the model paraphrasing beyond the source, and the quoted passage can itself be wrong, as the Dow Jones complaint describes. The check happens after the answer, in the source.
How do I use Perplexity's Academic focus, Deep Research and exports?
Open every citation, read the passage, and take the reference from the paper's own page rather than from Perplexity's text. What follows is from Perplexity's pages, checked on 22 September 2026.
- Opening sources. Each numbered citation is a link, and the sources panel under an answer lists them all. Hover over or select a source to see its domain label, if it has one.
- Academic focus. The Pro Search page lists Academic among the focuses you can choose, and the 2024 getting-started guide says it prioritises peer-reviewed journals and scholarly articles. No Perplexity page read for this article says which index it searches; the developer documentation restricts academic search by domain, to sites such as arxiv.org, pubmed.ncbi.nlm.nih.gov and semanticscholar.org, a filter on the web rather than a scholarly database. Organisation accounts cannot use focus mode.
- Deep Research. The report streams into a file you can edit and share, and a progress display shows which sources are being read. Its reference list needs the same check: Rao, Wong and Callison-Burch (2026) found that other makers' deep research agents cited more sources and invented URLs at higher rates than ordinary search-backed models.
- Exporting citations. Perplexity's pages do not describe exporting a reference list in a citation style. Build each reference from its DOI instead: EdCitation's Cite a source sets it in APA 7, MLA 9, Chicago author-date, Harvard, IEEE or Vancouver from the publisher's record.
How do I check a reference list that came from Perplexity?
Check each reference against the publisher's record, not against Perplexity, then check that the source says what your sentence says. For one reference, resolve the DOI, compare the title and authors on the page it opens, and read the passage; how to check whether a reference is real walks through it.
EdCitation's Verify references is the best tool for the first half of that job. Perplexity answers every question, "does this source exist?" included, with a web search and a model's sentence; Verify references never writes a reference and answers from the publisher's record instead. That matters more than Perplexity's working links suggest: in the one peer-reviewed test of its reference lists, Özbek and Bağcıer (2026), Perplexity had the worst hallucination score of three tools, 6.51 against ChatGPT's 1.81. Paste the list or upload the paper and each entry is sorted into verified, doubtful ("check this") or not found, with retracted papers flagged; an index that stays silent yields "could not check", never a false "not found". The second half, whether the page supports your sentence, stays with you. The check is free and needs no account; Pro, $8 a month, adds tools such as Mechanics QA for the paper's formatting, and Max, $24 a month, adds Theoretics QA and the Library.
When a source turns out to be a press release, search for the study it reports in Find sources and build the reference from the paper's DOI. If your course allows Perplexity, cite it as the guide to citing AI tools describes.
Quick questions
Does Perplexity make up references?
Rarely in the way a chatbot without search does, because Perplexity retrieves real pages first. Its failures are a real page that does not support the sentence, a summary cited instead of the paper, and an overstated finding; publishers have also alleged in court that it attributed invented text to them.
Is Perplexity more accurate than ChatGPT for citations?
It depends on the task. In the Tow Center's 2025 test of tracing news excerpts, Perplexity had the lowest error rate the report gives, 37%. In a 2026 test of manuscript reference lists, Perplexity had the worst hallucination score of the three tools and ChatGPT the best.
Does Perplexity's Academic focus search only peer-reviewed papers?
Perplexity says the Academic focus prioritises peer-reviewed journals and scholarly articles, but its pages do not name the index behind it, and its developer documentation implements academic search as a filter on web domains. Treat the results as web pages until you have opened each one. EdCitation's Find sources works the other way round, searching about 300 million published works rather than the web.
Are Perplexity Deep Research citations more reliable?
Deep Research reads and cites more sources, and more citations do not mean fewer wrong ones: a 2026 study of other makers' deep research agents found they invented URLs at higher rates than ordinary search-backed models. Check the list the same way.
References
- Dow Jones & Co. v. Perplexity AI, Inc., No. 1:24-cv-07984-KPF (S.D.N.Y. Jan. 28, 2025). Second amended complaint. https://business.cch.com/ipld/DowJonesCoPerplexityAISAC20250128.pdf
- Encyclopaedia Britannica, Inc. v. Perplexity AI, Inc., No. 1:25-cv-07546 (S.D.N.Y. Sept. 10, 2025). Complaint. https://www.susmangodfrey.com/wp-content/uploads/2025/09/Britannica-v.-Perplexity-Complaint.pdf
- Jaźwińska, K., & Chandrasekar, A. (2025, March 6). AI search has a citation problem. Columbia Journalism Review. https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
- Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines [Preprint]. arXiv. https://arxiv.org/abs/2304.09848
- Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W., Koh, P. W., Iyyer, M., Zettlemoyer, L., & Hajishirzi, H. (2023). FActScore: Fine-grained atomic evaluation of factual precision in long form text generation [Preprint]. arXiv. https://arxiv.org/abs/2305.14251
- Özbek, İ. C., & Bağcıer, F. (2026). Reference hallucination in AI-assisted academic writing: A comparative analysis of ChatGPT, Gemini, and Perplexity in rotator cuff literature. Indian Journal of Orthopaedics, 60(8), 1949-1956. https://doi.org/10.1007/s43465-026-01807-0
- Perplexity. (n.d.). Academic and scholarly search. Perplexity API documentation. https://docs.perplexity.ai/docs/cookbook/articles/academic-search/README
- Perplexity. (2024, October 1). Getting started with Perplexity. https://www.perplexity.ai/hub/blog/getting-started-with-perplexity
- Perplexity. (2026a, September 3). How does Perplexity work? Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work
- Perplexity. (2026b, September 9). Understanding source labels. Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/20260806-understanding-source-labels
- Perplexity. (2026c, September 9). What advanced AI models are included in my subscription? Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/10354919-what-advanced-ai-models-are-included-in-my-subscription
- Perplexity. (2026d, September 3). What is Pro Search? Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/10352903-what-is-pro-search
- Perplexity. (2026e, September 15). What's new in Advanced Deep Research. Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/13600190-what-s-new-in-advanced-deep-research
- Perplexity. (2026f, September 5). Which Perplexity subscription plan is right for you? Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/11187416-which-perplexity-subscription-plan-is-right-for-you
- Rao, D., Wong, E., & Callison-Burch, C. (2026). Detecting and correcting reference hallucinations in commercial LLMs and deep research agents [Preprint]. arXiv. https://arxiv.org/abs/2604.03173
- Roy, A. (2025). Perplexity [Product review]. Journal of the Canadian Health Libraries Association, 46(2), 47-52. https://doi.org/10.29173/jchla29877
- Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5