The Citation Problem
For the first two years of the generative AI boom, healthcare professionals had a serious problem: AI models invented sources. If you asked ChatGPT for clinical studies on a specific supplement protocol, it would invent studies, invent authors, and generate fake PubMed URLs.
This made general-purpose LLMs unusable as a citation source for clinical work. The real baseline was never Google Scholar. It was a structured PubMed or MEDLINE search with MeSH terms, filters, and a search string you can save and rerun.
In 2026, Perplexity AI narrowed this problem. It did not eliminate it. Perplexity's own Terms of Service still warn that the service may generate output containing incorrect, biased or incomplete information, and tell you not to rely on it for medical advice. Grounding an answer in retrieved sources changes the failure mode. It does not remove it.
How Perplexity Actually Works
Perplexity is an "answer engine." When you ask it a question, it does not just predict the next word based on its training data. Instead, it:
1. Breaks your query down into multiple search terms.
2. Scans the live internet for the most authoritative sources.
3. Reads those sources in real-time.
4. Synthesizes an answer.
5. Attaches numbered citations to the claims in its answer. Note the limit: citations are attached to an answer, not guaranteed on every sentence. Some sentences carry no marker at all, and one marker often covers several sentences. The unmarked sentences are exactly the ones you have to check by hand.
If Perplexity says, "Vitamin D supplementation has been shown to reduce the risk of acute respiratory tract infections," there will be a little `[1]` next to the sentence. Clicking that `[1]` takes you to whatever page it retrieved. Sometimes that is the PubMed record or the journal article. Often it is a news write-up, a press release, or a secondary summary of the study rather than the study. Perplexity says so itself: its source labels reflect the type or category of an overall website, not the accuracy of any single article on it, and a label is never a substitute for reading the source yourself.
That sentence is the example worth sitting with. If the citation lands on the 2017 individual participant data meta-analysis, the sentence is defensible for 2017. It is not the current picture: later aggregate data meta-analyses in Lancet Diabetes and Endocrinology, including a 2025 update across 46 trials and 64,086 participants, did not find a statistically significant reduction in acute respiratory infection risk. "Has been shown to" is doing work the evidence no longer supports. The citation was real, the source was real, and the sentence was still wrong.
The Pro Search Feature
Pro Search no longer stops to interview you. Earlier versions sometimes asked a clarifying question before searching; the current behavior is an automatic classifier that judges query complexity and plans the search itself. That shifts the work back onto you: put the population, the setting and the date window in the prompt yourself.
Academic Focus Mode
Perplexity lets you narrow where it looks. The control has moved and been renamed more than once. Today you set it through the source selector in the search bar rather than a standalone "Focus" menu, and on some organization accounts the manual control is gone entirely because Perplexity classifies the query for you. Check what your account actually shows before you assume a filter is applied. If you are writing a clinical protocol or a course for other practitioners, you can set the focus to "Academic."
In this mode Perplexity weights scholarly sources over blogs, news sites and consumer health portals like WebMD. It does not filter to peer-reviewed work only. Perplexity's own documentation puts preprint servers in the same academic bucket as journals, so medRxiv, bioRxiv and arXiv results come back looking exactly as authoritative as a published trial. They are not the same thing. A preprint has not cleared peer review, and the version you read may change or never publish.
The Failure Mode Citations Do Not Fix
A citation proves the source exists. It does not prove the sentence above it is a fair reading of that source.
The common failure is not a fake study any more. It is a real study, correctly linked, summarised in a way that quietly loses the population, the effect size, the confidence interval, or the hedge the authors put in the abstract. A trial in adults becomes a claim about everyone. A subgroup signal becomes a headline finding. A non-significant trend becomes "has been shown to".
That failure is invisible if you read the summary and trust the number next to it. It is obvious the moment you open the source. So for any claim you would put in a protocol, a patient handout or a course: open the citation, find the sentence in the source that supports the claim, and confirm the population and the effect size match. Perplexity says the same thing in its own documentation. Treat its output as a map to the literature, not as the literature.
What You May Not Type Into It
Perplexity's consumer and Pro tiers are not covered by a Business Associate Agreement. Perplexity's own Enterprise terms state that a customer may not use the services to process protected health information unless the customer and Perplexity have executed a BAA. A BAA is available only through an enterprise agreement, and only if you actually execute one. If you are a solo practitioner on a $20 subscription, you do not have one.
That means no patient identifiers, no chart excerpts, no lab panels tied to a person, and no case narrative specific enough to re-identify someone. HHS requires the business associate contract to be in place before you hand PHI to any vendor, not after. Perplexity is a literature tool. Ask it about the drug, the mechanism, the guideline and the population. Keep the patient in your EHR. This is general information, not medical advice or legal advice.
How to Use It
If you are writing patient education materials, building a course, or working through a complex case, use Perplexity to orient yourself and to find the names of the trials and guidelines you need. Then run the actual search in PubMed and read the primary literature. An answer engine reads the handful of pages it retrieved. It cannot tell you what it missed, and it cannot show you a search string you can reproduce or defend. For anything that touches a licensed decision, the AI search is the on-ramp to the literature search, not the substitute for it. Perplexity is a strong general-purpose answer engine with attribution built in. It is not the only one, and it is not purpose-built for clinical evidence. For a question that will touch patient care, a tool built on the biomedical literature, or a straight PubMed search you run yourself, will beat a general web answer engine on completeness and on knowing what it missed.