Perplexity search operators and prompts
Perplexity searches before it writes, so a good prompt does two jobs: it decides what gets read, and then what to do with it. The operators are the first half, and most people never use them.
It is a search engine that writes
Most prompting advice assumes the model answers from what it already knows. Perplexity does not: it runs a search, reads what comes back, and writes an answer with citations attached. That one difference changes what a good prompt looks like. You are not describing a task to an assistant — you are describing which documents should be found, and then what to do with them.
So the two halves of a strong Perplexity prompt are a retrieval half (the words and operators that decide what gets read) and an instruction half (what to do with it). Most weak prompts have only the second.
The search operators worth knowing
Perplexity passes a handful of classic search operators through to its retrieval step. They go anywhere in the query, and they are the fastest way to control what it reads.
site:— restrict to one domain.site:nih.gov,site:arxiv.org,site:gov.uk. The single most useful one, because it is how you say "primary sources only" in four characters.-site:— exclude a domain. Useful for cutting out an aggregator that dominates a topic, or a competitor's blog when you want independent coverage.filetype:— restrict by file type.filetype:pdfis the one that earns its keep: reports, papers and filings are PDFs, and blog summaries of them are not.after:andbefore:— bound the dates.after:2024is how you stop a fast-moving topic being answered from a well-ranked page written three years ago.inurl:andintitle:— require a word in the URL or the page title. Narrow, but excellent for finding documentation and changelogs.OR— widen deliberately, in capitals."semaglutide" OR "tirzepatide"catches both names for a topic that is written about under either.
These eight are the chip row that appears above the keys when the target is set to Perplexity — site:, -site:, filetype:pdf, after:, before:, inurl:, intitle: and OR. Nobody remembers all of them, which is exactly why they are one tap rather than one more thing to recall.
Shape the question, not the topic
A topic gets you an encyclopedia entry. A question gets you an answer. The difference is usually one clause: what decision the answer is for, or what you already know and do not need repeated.
research on intermittent fasting
Summarize peer-reviewed research published since 2024 on intermittent fasting and metabolic health. Prioritize systematic reviews and meta-analyses, cite every source, and flag where the findings conflict. filetype:pdf after:2024 site:nih.gov
Three clauses do most of the work in that rewrite: a date bound, a preference for the kind of evidence that settles arguments, and an explicit instruction to surface disagreement rather than smooth it over. The last is the one people leave out, and it is the one that turns a summary into research.
Ask what would change the answer
The failure mode of an AI search tool is a fluent, well-cited consensus that does not exist. Two instructions defend against it, and both are worth pasting into any serious research question:
- "Where sources disagree, say so and name which sources take which side."
- "If the evidence is thin or mostly from one group, say that rather than answering confidently."
Neither costs you anything when the consensus is real, and both save you when it is not.
Use follow-ups instead of longer prompts
Perplexity keeps the thread, so the second question is where the value is. Rather than writing one enormous prompt, ask a narrow question, look at what it cited, and then push on the weak part: "the second source is a press release — find independent coverage of the same result", or "restrict that to the UK". You are steering a search, and steering works better in small corrections than in one long instruction written before you knew what would come back.
Read the citations, not just the answer
Every claim comes with a numbered source, and the habit worth building is checking one or two of them on anything that matters — particularly the date. A correctly cited page can still be years out of date, and an answer that reads as current because it is well written is the exact thing the after: operator exists to prevent.
The operators, one tap away
Set the target to Perplexity in Prompt AI Keyboard and the eight operators on this page sit in a row above the keys — tap to insert, type the domain after it. Tap Enhance and your rough question is rebuilt as a research prompt, wherever you are typing it.
FAQ
What search operators does Perplexity support?
The classic ones carry through to its retrieval step: site: and -site: to include or exclude a domain, filetype: (filetype:pdf especially) to restrict by document type, after: and before: to bound the dates, inurl: and intitle: to require a word in the URL or title, and a capitalised OR to widen a query deliberately.
How is prompting Perplexity different from prompting ChatGPT?
Perplexity searches before it writes, so half of a good prompt is about which documents get found — that is what the operators are for — and half is what to do with them. A prompt that only describes the task, with nothing steering retrieval, leaves the most useful control unused.
How do I stop Perplexity giving me a confident answer from weak sources?
Bound the dates with after:, prefer primary sources with site: and filetype:pdf, and add two instructions: say where sources disagree and name which takes which side, and say when the evidence is thin rather than answering confidently. Then check a citation or two, especially its date.