How Accurate Are Perplexity's Citations? What the Research Actually Shows

TL;DR: No one has published a fresh count of how often Perplexity's citations are accurate. The closest real measurement is a March 2025 Columbia Journalism Review study that found Perplexity wrong on 37 percent of one specific task, better than ChatGPT Search's 67 percent on the same test. A citation existing is not the same as a claim being checked; here is what the research supports and how to verify a citation in under two minutes.

You will find pages claiming to have run their own hundred-citation audit of Perplexity and returned a tidy accuracy percentage. We looked for the underlying methodology behind claims like that and could not find one that would survive a second read: no published list of the citations checked, no scoring rubric, no access date. A number with no method behind it is not research, it is a number that sounds like research.

So this post does something less flattering and more useful. It reports the one study that actually ran a citation-accuracy test with a published methodology, scopes its findings honestly (including a figure that makes Perplexity look better than the tool people usually compare it to), and then hands you the two-minute habit that matters more than any percentage: how to check a citation yourself before you repeat what it says.

What does the actual research say about Perplexity's citation accuracy?

The relevant study is the Tow Center for Digital Journalism's "AI Search Has a Citation Problem", published in Columbia Journalism Review on March 6, 2025, by Klaudia Jaźwińska and Aisvarya Chandrasekar. It is worth reading in full rather than trusting a summary of it, including this one.

The methodology: the researchers selected 20 news publishers with varying crawler-access policies, pulled ten articles from each, and hand-picked direct excerpts chosen so that pasting them into Google would surface the original source in the first three results. They fed each excerpt to eight tools, OpenAI's ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, xAI's Grok-2 and Grok-3, and Google's Gemini, and asked each one to identify the article's headline, publisher, publication date and URL. That is 20 publishers times 10 articles times 8 chatbots, or 1,600 queries total, exactly as the study states it.

The headline result, quoted directly: "the chatbots often failed to retrieve the correct articles. Collectively, they provided incorrect answers to more than 60 percent of queries. Across different platforms, the level of inaccuracy varied, with Perplexity answering 37 percent of the queries incorrectly, while Grok 3 had a much higher error rate, answering 94 percent of the queries incorrectly."

Two things in that table matter more than the headline number. First, ChatGPT Search performed worse than Perplexity on the exact task being compared, 134 of its 200 responses misidentified the article, a 67 percent error rate. Anyone citing this study to single out Perplexity as uniquely unreliable is leaving out the tool that did worse on the same test. Second, the study's confidence finding applies broadly. Most tools rarely used hedging language like "it appears," acknowledging a knowledge gap only occasionally: "I couldn’t locate the exact article." ChatGPT specifically signaled a lack of confidence in only 15 of its 200 responses despite getting 134 of them wrong.

One more scoping note the study makes about itself, worth carrying forward: premium tiers (Perplexity Pro, Grok 3) answered more questions correctly than their free equivalents but also produced more wrong answers overall, because they declined to answer less often.

No comparable study has re-run this test since March 2025. If you read a newer number without a named methodology behind it, that is the tell.

Why isn't a correct-looking citation always the real source?

Two more findings from the same study are worth carrying forward, because they're a different failure mode from a wrong headline, and neither one is fixed by a higher accuracy percentage.

The first is syndication standing in for the original. The study found that despite an active partnership between Perplexity and the Texas Tribune, "Perplexity Pro cited syndicated versions of Tribune articles for three out of the ten queries, while Perplexity cited an unofficial republished version for one." A citation here points to a real, working page. It just isn't the page the publisher wrote, which matters if you're trying to credit or contact the original source, and matters more if a syndicated copy has been edited, cut, or is simply older than the version you actually need.

The second is a crawler-access question that looks, on the surface, like an accuracy win. The study reports that Perplexity's free version "correctly identified all ten excerpts from paywalled articles we shared from National Geographic", despite National Geographic having disallowed Perplexity's crawlers and having no formal relationship with the company. That should not have been possible if Perplexity's retrieval respects the access National Geographic's robots.txt file requests. The study's own read on it, phrased carefully around a direct quote of Perplexity's own claim: Perplexity, despite claiming that it "respects robots.txt directives", may have disregarded that publisher's crawler preferences. A correct answer built on content a publisher tried to block is not evidence the retrieval is trustworthy; if anything, it is evidence the access rules and the citation behavior aren't as tightly coupled as they should be.

Neither of these shows up in a bare accuracy percentage. Both show up the moment you open the citation and check where it actually points.

Does a citation mean the claim was checked?

No, and this is the part that outlasts any single study's numbers. A citation records that a page was retrieved and used to generate part of an answer. It does not record that the sentence above it accurately represents what that page says, that the page is current, or that the page even supports the specific number attached to it.

Perplexity's own help centre is direct about this. Its source-labels article, last updated August 7, 2026, explains a shield icon that marks certain domains as Government, Academic, or Trusted, based on criteria like whether a site corrects its mistakes and separates news from advertising. Then it states the limit of what that means: "A label describes the website as a whole, not any single article or claim on it. It is not an opinion about the site’s views. A label can change as sites are reviewed again. And a label is never a substitute for reading the source yourself."

Its closing line is the one worth remembering longer than any accuracy percentage: "Perplexity aims for accuracy, but reviewing the original sources yourself remains the best way to build confidence in an answer." That is Perplexity's own guidance, not a critique of the product. A shield icon on a domain tells you the website has a track record. It tells you nothing about the one sentence you are about to repeat in a meeting.

The surfaces changed since most guides about this were written

Anyone citing Perplexity's product surfaces should check them fresh, because two changed quietly enough that older guides are now describing things that do not exist any more.

Perplexity Labs has no help-centre article. Checking the current sitemap directly turns up 2,413 URLs, 207 of them English-language articles, and zero of them mention Labs by name. That is not proof the name never appears anywhere in the product, but it means Perplexity is not documenting it as a supported feature today, and a guide instructing you to open it is pointing at something outside the current help centre.

Spaces was renamed to Projects. The old help-centre URL for the Spaces article still returns a normal 200 response rather than redirecting or erroring, which is exactly why checking only the status code would miss this: the page it now serves is titled "What are Projects? | Perplexity Help Center", and describes Projects, a persistent workspace with roles, custom instructions and connected sources, throughout. A status-code check says the page works. Only reading the title tells you it is not the page you asked for.

That second point generalizes past Perplexity. A 200 response is not evidence you retrieved the content you expected, it is evidence a server answered. The only way to know you have the right page is to read what came back and check its own title against what you asked for.

Our own site doesn't agree with itself on this either

In the interest of not applying a standard to Perplexity that we dodge ourselves: our FAQ page currently gives three different answers about whether we support Perplexity.

One answer states: "Prompt Architects optimizes prompts for the latest models from OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), xAI (Grok), Perplexity, and more." That reads as current support. A second answer, in the same FAQ section, says: "Upcoming integrations include Perplexity, Midjourney, and more." That reads as not yet shipped. A third answer, in the Global Variables section, states variables "work across every integration we support" and then lists Perplexity by name in that list. Our /integrations page marks Perplexity's status as Available, which agrees with the first and third answers, not the second.

We are reporting this rather than picking a side, because picking a side here is exactly the failure mode this whole post is about: it is faster to assert an answer than to check one, and the FAQ page currently has an answer that needs fixing rather than restating.

The two-minute method that matters more than any accuracy number

Whatever the current percentage is, and there is no reliable fresh one, the habit that actually protects you is unglamorous:

1. Find the claim you would be embarrassed to be wrong about. Not every sentence in an answer needs checking; the one with a number, a date, or a name in it does.
2. Open the specific citation attached to that sentence, not just the top of the source list.
3. Read the paragraph the claim is drawn from, not just the headline. The CJR study's central finding was tools misattributing sources and citing syndicated copies instead of originals; both of those are visible the moment you open the page.
4. Check whether the source is the original or a syndicated copy. The study found Perplexity Pro citing syndicated Texas Tribune articles for three of ten queries despite an active partnership with the original publisher.
5. If the sentence still says what you thought after steps 2 to 4, use it. If it doesn't, the citation existed and the claim was still wrong.

Before you repeat a cited claim, paste this into the same chat:

"Open the source you just cited for [claim]. Quote the exact sentence
that supports it. If the source doesn't contain a sentence that
supports the claim as stated, say so directly rather than restating
the claim."

That prompt does not make the underlying retrieval more accurate. It forces the one step the Tow Center study found chatbots skip almost universally: admitting when a source does not actually say what the answer implies. If you want prompts built for Perplexity's research modes specifically, our Perplexity prompt templates and the guide to writing search-shaped prompts go further into structure; this post is only about the verification step that comes after the answer arrives. For the mode Perplexity actually recommends for a citation-heavy report, our Focus-modes guide covers Deep Research and Model Council in detail.

What will make this post stale first

In order of how quickly I expect it to happen:

1. A newer citation-accuracy study, if one gets published with a real methodology. Until then, March 2025 is the only dated, named measurement available, and it is now a year and a half old.
2. Perplexity's product surfaces, again. Projects gained roles and connected-tools settings recently enough that the article documenting them is still dated within the last month.
3. Our own FAQ contradiction, which should get fixed rather than stay a live example in this post.

Where a prompt tool fits, honestly

It doesn't fit into the retrieval. Prompt Architects does not crawl the web, does not rank sources, and cannot make a citation more accurate; Perplexity's retrieval pipeline is Perplexity's, and no browser extension changes what it fetches or how it fact-checks itself.

What a saved, reusable prompt does is make the verification habit above easier to actually run every time, rather than the first time you remember to. The free plan includes 5 prompt enhancements per day, forever, according to our own FAQ page, which is enough to test whether keeping a source-checking prompt on hand changes how often you actually open the citation.