Perplexity Deep Research
Give it a complex question, walk away, come back to a cited mini-paper.
Your manager just asked you to write a one-pager comparing four vendors in a market you do not know. It is due Thursday. A regular Google search will not get you there, and you do not have the time to open thirty tabs.
Why this tool matters
Perplexity Deep Research is an AI agent that performs a ten-to-twenty-minute research session on your behalf, opens dozens of web pages, reads them, takes notes, and produces a cited multi-page report. It is the first tool in this course that is genuinely an agent in the useful sense: you set it loose on a goal and it works while you do other things.
Regular Perplexity (the Ask view) is a search-augmented chatbot: you type a question, it searches the web, synthesizes an answer with citations. Deep Research is a different mode. You give it a research goal (“compare the regulatory landscape for stablecoins in the US, EU, and Singapore in 2026”), and it runs for 10–20 minutes. The output is typically 2,000–5,000 words, with 30–70 citations, structured as a report.
For any research task where breadth matters more than a single right answer — competitive analyses, literature scans, regulatory overviews, due diligence, background briefings for meetings — Deep Research replaces what would have been a day of work with roughly twenty minutes of machine time.
Setup
Account: free Perplexity account at perplexity.ai. The free plan gives limited Deep Research runs per day; the $20/month Pro plan removes the ceiling and unlocks higher-quality model choices (Claude, GPT, and Perplexity's own Sonar).
Prerequisite: a specific, somewhat complex question. Deep Research wastes its power on simple factual questions that regular Perplexity handles better.
Walkthrough
Step 1: Pick Deep Research mode
Open perplexity.ai. In the query box, click the mode selector and pick Deep Research. The affordance tells you this is going to take longer than a normal query — do not fight it.
Step 2: Write a genuinely complex question
The quality of Deep Research output depends almost entirely on question quality. Bad: tell me about stablecoins. Good: What is the current regulatory treatment of fiat-backed stablecoins in the US, EU, and Singapore as of 2026? Compare reserve requirements, issuance licensing, and consumer protections. Cite authoritative sources.
Step 3: Submit and get out of the way
Perplexity begins to stream its progress — you will see it opening sites, reading, and taking notes. You can watch this if you want (it is mesmerizing the first few times), but you can also close the tab. Results wait for you.
Step 4: Read with skepticism, then click through
When the report lands, read it as a draft — not gospel. Every paragraph should have citations. For any claim that feels important or surprising, click the citation and read the actual source. Deep Research is strong, but it still occasionally misreads a nuanced source.
Step 5: Follow up to tighten
At the bottom of the report, you can continue the conversation. Ask: Of the three jurisdictions, where are reserve requirements strictest? Cite the specific regulation. This turns Deep Research into an iterative process rather than a one-shot lottery.
Step 6: Export the artifact
Use the Export as PDF or Copy as Markdown option in the menu. You now have a document you can edit into your own brief, with an audit trail of every source.
Your turn
Basic: A comparative briefing
Pick three products, tools, frameworks, or policies in your field. Ask Deep Research to produce a comparison briefing on a specific dimension — pricing, feature set, adoption, regulatory posture, whatever matters in your world. Read the report. Identify one claim you did not already know, and click through to the source to verify it.
Advanced: A real due-diligence memo
Pick a real decision you or your team is currently weighing — a vendor, a hire, an investment, a methodology choice. Write a 3-paragraph Deep Research prompt that clearly states the decision, the dimensions you care about, and the level of rigor you need in the sources.
Run it. Then spend 15 minutes doing a citation audit: open ten random citations and verify the report characterizes each source correctly. Flag any mischaracterizations. Edit the report into a 500-word memo you would be comfortable forwarding.
The skill this exercise is building is not AI literacy — it is AI output literacy: knowing exactly how much to trust a machine-generated report.
Pitfalls and pro tips
Deep Research is not peer review. It reads the open web. Low-quality sources (marketing pages, SEO-farm content) can end up cited alongside authoritative ones. Always check citations for anything decision-critical.
Recency matters. For fast-moving topics (policy, pricing, product releases), include the year in your prompt and explicitly ask for recent sources. Otherwise it will sometimes anchor on older articles.
One question per run. Deep Research performs worse when the prompt bundles multiple unrelated questions. Split them into separate runs.
How it compares
Deep Research competes with ChatGPT Deep Research, Gemini Deep Research, and Claude Research — all released in a tight window in late 2024 and early 2025. At the current state of the art, they are roughly comparable on easy prompts and noticeably different on hard ones. Perplexity is usually strongest on recent events and comparison structures; ChatGPT is strongest when the output needs to be a long, well-written essay; Gemini is strongest when the topic is in Google's knowledge graph. For a professional brief, try the same prompt on two of them and compare.
When to use — and when not to
Use Deep Research when you have a complex, multi-sourced question and the output needs citations a reviewer could click. Competitive analyses, regulatory scans, literature reviews, background briefings — all ideal.
Do not use Deep Research when the answer is in one known source (use NotebookLM), when you need a single factual answer (regular Perplexity is five seconds), or when the topic is so specialized that web sources are insufficient (you need expert interviews, not AI).