Written and maintained by CASRAI Editorial Board
Last updated
Perplexity AI shows up constantly in “AI tools for research” lists alongside Elicit, Consensus, and Semantic Scholar, but it was not built as an academic literature-discovery tool. It is a general-purpose AI answer engine: type a question, and it searches the open web, then returns a synthesized answer with inline citations to whatever pages it retrieved — news sites, vendor pages, forum threads, and peer-reviewed papers mixed together, with no filter distinguishing one from another in the interface. That distinction matters more than it sounds, because it changes what a research office should and should not trust Perplexity to do.
What Perplexity actually is
Perplexity was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski. Its free tier requires no registration; a paid Pro tier adds document search, API access, and the ability to choose which underlying model answers a query. Perplexity built its own default model, Sonar (based on Meta’s Llama 3.3), and a second in-house model, R1 1776 (based on DeepSeek R1); Pro subscribers can also route a query to GPT-5.4, Claude 4.6, or Gemini 3.1 Pro instead. None of this is academic-specific — it is the same infrastructure whether the query is “best restaurants near me” or “effect sizes in meta-analysis.”
Perplexity Deep Research: what the mode is for
Perplexity offers a separate “Deep Research” mode, distinct from a standard query, intended for questions that need more than one search pass — it runs a multi-step research process and assembles a longer, more structured report with citations, rather than a short synthesized answer. Vendor-published specifics on exactly how many sources or search passes a given Deep Research run performs change frequently enough, and are marketing-adjacent enough, that this guide does not repeat a number here without being able to independently verify it against the live product; check Perplexity’s own current documentation if an exact figure matters for your use case. What is safe to say: Deep Research is built for breadth-first scoping across a live, mixed web index, not for a bounded, reproducible search of the peer-reviewed literature the way a systematic-review tool is.
Why citation reliability is the load-bearing question for research use
For casual use, an AI answer engine’s occasional citation error is a minor annoyance. For academic work — where a claim’s credibility depends on it tracing to a real, checkable source — it is the central risk, and Perplexity has a documented history of it. Dow Jones and the New York Post filed suit against Perplexity in 2024 alleging, among other claims, that Perplexity had harmed their brands “by attributing hallucinated quotes, for example, on F-16 jets for Ukraine, to articles that did not include them” — a citation fabricated a quote, then attributed it to a specific real publication that never printed it. Separately, developer Robb Knight and Wired reported in June 2024 that Perplexity was crawling sites using undisclosed IP addresses and spoofed user-agent strings while ignoring robots.txt directives, despite public claims to the contrary; Cloudflare published research in August 2025 corroborating that Perplexity used “stealth” crawlers to bypass site-level blocking. Forbes (June 2024), the New York Times (October 2024, cease-and-desist), the BBC (June 2025), Japan’s Yomiuri Shimbun (August 2025), the Asahi Shimbun and Nikkei (August 2025), and Reddit (October 2025) have all pursued copyright or scraping claims against the company on related grounds.
None of this means Perplexity’s output is generally unreliable for everyday use — it means the specific failure mode (a confident-sounding citation that does not actually say what it is credited with saying) is a documented, recurring pattern, not a hypothetical edge case, and is exactly the failure mode that matters most when a citation is headed into a manuscript, grant narrative, or literature review rather than a casual search.
How Perplexity compares to purpose-built academic discovery tools
Tools built specifically for the research workflow constrain their search space and their claims in ways a general answer engine does not:
- Elicit and Consensus search a corpus of academic papers specifically and extract structured findings (sample sizes, effect directions, study design) rather than freeform web text — see CASRAI’s direct Perplexity vs. Consensus comparison for a side-by-side on research-question fit.
- Semantic Scholar (Allen Institute for AI) indexes a large open corpus of scholarly papers and layers AI-generated summaries and influence metrics on top of citation data specifically, not the general web.
- Scite classifies each citation to a paper by whether the citing work supports, contrasts with, or merely mentions the claim — a structured signal no general answer engine currently provides.
- Connected Papers and Research Rabbit build visual citation-network maps from a seed paper, useful for mapping a field’s structure rather than answering a single question.
The practical difference: a tool scoped to the academic literature can tell you what it searched and what it left out. A general answer engine searching the open web cannot make that same guarantee, because “the web” is not a bounded, versioned corpus the way a paper database is.
Practical guidance for research use
- Use Perplexity (including Deep Research) for early-stage scoping — getting oriented on an unfamiliar topic, finding terminology, or generating search terms to run against a proper academic database — not as the literature search itself.
- Treat every citation Perplexity surfaces as a lead, not a fact: open the source it cites and confirm the claim is actually there before it goes into any manuscript, grant narrative, IRB submission, or policy document.
- Do not cite Perplexity’s synthesized answer as a source in academic writing; cite the underlying primary source it points to, once you have verified it says what Perplexity says it says.
- For a bounded, reproducible, citation-classified search of the peer-reviewed literature, prefer a tool built for that job — see the comparisons above.
Frequently asked questions
Is Perplexity AI accurate enough to cite in a research paper?
Not directly. Perplexity is a synthesis layer over the open web, and it has a documented, litigated history of misattributed and fabricated citations (see above). Treat its output as a starting point for locating sources, then cite and verify the primary source itself — never Perplexity’s summary of it.
What is Perplexity Deep Research, specifically?
A slower, multi-step mode built for broader questions: instead of one search pass and a short answer, it runs an extended research process and produces a longer, structured, cited report. It searches the general web, the same way standard Perplexity queries do — it is not a bounded search of the peer-reviewed literature.
Is Perplexity free to use for research?
Yes, a free tier is available without registration. A paid Pro tier adds document search, API access, and the ability to select a different underlying model (including GPT-5.4, Claude 4.6, or Gemini 3.1 Pro) instead of Perplexity’s own Sonar model.
How is Perplexity different from Elicit or Consensus for a literature review?
Elicit and Consensus search a corpus of academic papers specifically and extract structured findings; Perplexity searches the general web and does not distinguish peer-reviewed sources from any other page type in its interface. For a systematic or even semi-structured literature review, a tool scoped to academic content is the better fit — see CASRAI’s Perplexity vs. Consensus comparison.
