Perplexity review: research quality and limitations

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of Perplexity. What follows rests on official documentation, published pricing and independent studies we read in full, all accessed on 16 October 2026. A tested update will follow, drawing on our citation test, Can AI research assistants substantiate their claims?, and our head-to-head Perplexity vs. ChatGPT for research.

Perplexity calls itself an answer engine. You ask a question, it searches the web and writes a reply with numbered links to its sources. That design is why researchers like it: every claim seems to come with a receipt. It is also why the right question about Perplexity is not "Does it cite?" but "Do its citations hold up, and when should you rely on them?"

This review answers that from the record. It covers what each plan costs and what each mode does, where answers draw their sources from, and how citations are shown. It also covers the terms for company data, the limits Perplexity itself documents, and what independent researchers found. It ends with the publisher lawsuits, reported only where we could check the court record. We give no scores: that waits for our own tests.

What Perplexity is, and what each mode is for

Perplexity is one product with several ways to ask a question, each going deeper than the last. According to its help centre:

• Search (Best mode) is the default. Perplexity picks a model for the question and returns a short cited answer. On Pro, Best mode has no quota.[1]
• Pro Search runs several searches and draws on "articles, academic papers, forums, videos, and more", depending on the focus you choose. Paid users can pick the AI model behind it.[2]
• Research (also called Deep Research) "performs dozens of searches, reads hundreds of sources", then writes a report you can export to PDF or a document. You cannot choose its model.[3][4] The help pages give slightly different run times (2 to 4 minutes on one page, around 4 to 5 on another), so expect a few minutes.
• Create files and apps, formerly Perplexity Labs, builds reports, spreadsheets, dashboards and simple web apps from research.[5][4]
• Comet is Perplexity's Chromium-based web browser. Its Comet Assistant is a browser agent that can act on web pages for you.[6][7]

Perplexity's support pages also say Pro answers carry "10x as many citations per answer" as free ones.[1] That is a vendor claim about quantity, not about whether the sources support the answer.

AI Chronicle

Plans and prices (as of 16 October 2026)

Sources: Perplexity's plan comparison, pricing page, Max help page and its October 2026 pricing explainer.[8][9][7][10] Perplexity does not publish exact Pro Search or Research numbers for Pro and Max. It also warns that advanced-model access "may be limited during weeks of especially heavy usage."[1] If you rely on volume, a published Enterprise allowance is easier to plan around than "average use".

Where its answers come from

By default Perplexity searches the open web in real time.[4] Pro Search lets you narrow the focus, for example to Academic, Finance or your own files.[2] Inside an Enterprise organisation the old focus menu is replaced by a "Choose sources" control: Web, organisation files or none. The organisation file repository was retired on 1 October 2026 in favour of Projects and file connectors.[11][8]

Two further source types matter for professional work. Premium sources such as Wiley, Statista, PitchBook and CB Insights are built in with daily caps, and Pro, Max and Enterprise users can connect licensed data they already pay for, such as FactSet or Morningstar.[12] And Perplexity's crawler policy says it will not index the text of a page blocked by robots.txt, but "may still index the domain, headline, and a brief factual summary."[13]

The practical point: the same question can draw on very different pools depending on mode and settings. Note which sources were switched on before you judge an answer, a habit our RAG explainer explains.

How it shows citations

Every answer carries numbered citations that link to the original pages.[4] Some cited domains now carry a shield icon with a Government, Academic or Trusted label, set by Perplexity's source review. The help page is careful about what a label means: it rates a whole website, not the page or claim in front of you, and "no label" is not a negative judgment.[14]

"[A] label is never a substitute for reading the source yourself."

— Perplexity Help Center, Understanding source labels (2026).[14]

Projects (formerly Spaces)

Spaces became Projects on 30 July 2026, and existing Spaces migrated automatically.[15] A Project holds searches, Perplexity Computer tasks, files and standing instructions, so a team can keep returning to one body of research. The documented limits:[16]

• Private by default, and Projects "cannot be made publicly queryable".
• Up to 5 contributors, or 9,999 if the owner belongs to an organisation or is a verified student or educator.
• Up to 10,000 files and 5 GB per Project, 100 MB per file, on every plan.
• Files added from a connector are one-time copies, so they will not update when the original does.

That last point is easy to miss: a report built in a Project can rest on a stale copy of a document someone has since revised.

Data terms for companies

Perplexity draws a sharp line between consumer and Enterprise accounts.[17][18]

• Free, Pro and Max: AI data retention is on by default, and "some data may be used to train AI models" unless you switch it off in settings. Opting out applies only to data collected afterwards, and previously collected training data cannot be deleted.
• Enterprise Pro and Max: data is "never used for AI training". Files attached to sessions are deleted after seven days. Organisations with 50 or more seats, or at least one Enterprise Max seat, can set custom session retention.[19]
• All plans: Perplexity says its agreements with providers such as OpenAI and Anthropic stop them training on Perplexity data.[20] Enterprise Pro is SOC 2 Type II certified.[18]

If staff use personal Pro accounts for work, those chats fall under consumer terms. That is a governance question, not a feature question.

What independent research found

Two large studies checked Perplexity's sources from the outside. Both used news questions in 2025, before the current Deep Research upgrade, so treat them as evidence about the pattern, not a verdict on today's product.

Tow Center, March 2025. Researchers gave eight AI search tools 200 news excerpts each and asked for the headline, publisher, date and URL. Perplexity answered 37% incorrectly; the worst tool, Grok 3, reached 94%.[21] But Perplexity Pro, while answering more prompts correctly than the free version, also gave more confidently wrong answers rather than declining. It correctly identified nearly a third of 90 excerpts from publishers that block its crawler, and the free version identified all ten paywalled National Geographic excerpts. The authors say this suggests Perplexity "may have disregarded" the publisher's crawler preferences. Perplexity Pro also cited syndicated copies of Texas Tribune articles in three of ten cases despite a partnership. Each excerpt was queried once.[21]

EBU and BBC, October 2025. Journalists from 22 public service media organisations assessed free versions of four assistants, with Perplexity giving 681 core answers. Perplexity had significant issues in 30% of responses, against 76% for Gemini, and significant sourcing issues in 15%.[22] Evaluators "often felt Perplexity was strong on sourcing". Yet they also logged irrelevant links, such as nine VRT sources including unrelated stories, and long source blocks the answer never used.

"Perplexity providing long lists of URLs without actually referring to them in the answers."

— NRK evaluator, quoted in EBU and BBC, News Integrity in AI Assistants (October 2025).[22]

Read together: Perplexity's sourcing compares well with rival assistants, but well is not reliable. More citations can look like more evidence while adding more to check. Our AI hallucinations explainer covers why fluent, cited answers still fail, and our citation test will measure Deep Research directly.

Publisher disputes and lawsuits

Perplexity has licensing deals with some publishers, including Gannett, TIME, Le Monde and Der Spiegel.[23] Others have sued. Only cases we confirmed in court records or primary reporting are listed, with their status as of 16 October 2026. None has produced a final ruling on whether Perplexity's copying is lawful.

One non-publisher case bears on Comet. Amazon sued over Comet Assistant shopping on Amazon.com. A preliminary injunction granted in March 2026 was vacated by the Ninth Circuit on 4 August 2026, which held the user, not Perplexity, "accessed" Amazon's computers; the case continues.[29][30] The court said its opinion does not establish a new legal regime governing agentic AI. Our explainer on AI training and copyright sets out why search-and-answer copying raises different questions from training.

Documented limits and risks

• Citation is not support. Both independent studies found cited sources that were wrong, irrelevant or unused.[21][22]
• Confident errors. Tow found the paid tier more likely to answer wrongly than to decline.[21]
• Unpublished consumer limits. Pro and Max allowances are described, not numbered, and advanced models can be throttled in heavy weeks.[8][1]
• Training by default. Consumer data trains models unless you opt out, and past data cannot be withdrawn.[17]
• Stale Project files. Connector imports are snapshots.[16]
• Open legal questions. The outcomes of the publisher suits could change what content Perplexity can show.
• Agent risk in Comet. An assistant that acts in your browser needs the same caution as any delegated access.

Who it suits, and who should skip it

It suits people who need fast, current answers with visible sources: analysts scanning a market, journalists finding leads, students starting a topic. It also suits teams that can buy Enterprise for its no-training terms and published allowances.

Skip it, or use it only as a lead generator, if you need a complete literature search (a scholarly tool such as those in our research tools shortlist searches defined paper indexes), or if your organisation cannot accept consumer data terms and won't pay for Enterprise. Skip it too if your work needs every claim verified and you lack time to open the sources.

How to test it yourself

1. Pick ten questions you can already answer from trusted sources, mixing recent news, a statistic and a niche topic.
2. Run each in Search, Pro Search and Research, noting which sources and focus were switched on.
3. Open every citation that carries a claim. Does the page load, and is it the original rather than a syndicated copy?
4. Find the sentence. Search the source for the specific claim, not just the topic.
5. Count unused sources in Research reports: listed links the text never relies on.
6. Rerun two questions a day later to see how much answers and sources change.
7. Check your data setting before testing with anything confidential.

A worked example (hypothetical)

This scenario is invented to show the workflow. No tool was run, and no outputs are reported.

Sam, a policy officer at a city council, must brief a committee on e-scooter safety by Friday. She runs a Research report in Perplexity and gets a tidy summary with dozens of numbered sources.

She treats it as a reading list. For the four statistics she plans to quote, she opens each citation and finds that one figure comes from a vendor blog summarising a study, not the study itself. She traces the original and replaces the link. A second source is a syndicated copy, so she cites the publisher. She saves the verified documents to a Project, noting that the files are copies and need rechecking before the next meeting.

The test she applies: could a committee member click any number and land on the passage it came from?

Questions readers commonly ask

Is Perplexity accurate?

On the published evidence, more accurate on sourcing than some rivals but far from error-free: 37% wrong in Tow's citation test and 30% with significant issues in the EBU/BBC study, both in 2025.[21][22] We have not measured the current version ourselves.

Is Pro worth it over the free plan?

For occasional questions, the free plan's unlimited basic search and 3 daily Pro Searches may be enough. Pro adds model choice, more Pro Searches and Research reports, but its limits are not numbered.[8]

Does Perplexity train on my searches?

On Free, Pro and Max, yes by default, until you switch off AI data retention; earlier data stays in training. Enterprise data is never used.[17]

How does it compare with ChatGPT for research?

That is the subject of our Perplexity vs. ChatGPT comparison. Before relying on either, work through how to verify AI-generated research.

The next step

If you are choosing between the two best-known cited research assistants, read Perplexity vs. ChatGPT for research next.

Sources and evidence note

Product descriptions, prices, limits and data terms come from Perplexity's help centre, pricing page, blog and changelog, and describe intended behaviour, not tested performance. Prices are US list prices as displayed on 16 October 2026. Independent evidence comes from the Tow Center (Columbia Journalism Review) article and the full EBU/BBC report. Case statuses come from court dockets and opinions, CNN and a Sankei report republished by ITmedia. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• Claude review: research and writing workflows
• What is RAG? Retrieval-augmented generation explained