Perplexity

Perplexity is an answer engine rather than a chatbot or a search engine. Ask a question and it runs the searches, reads the results, and writes a synthesized answer with a numbered citation on each claim. The citations are the product: you can click through and check the sentence you're about to repeat in a meeting, which is exactly the affordance a general assistant with a browsing tool doesn't reliably give you.

Why "answer engine" is a real distinction

A search engine returns ten documents and makes you do the synthesis. A chatbot synthesizes from memory and may confabulate. Perplexity does retrieval first and generation second, and it surfaces the retrieval — so its failure mode is "cited a weak source" (visible, checkable) rather than "invented a fact" (invisible until it bites you). That does not make it infallible; it makes it auditable, which is the more useful property for research work.

The modes, and when to use each

• Search — the default. Fast, a handful of sources, good for "what is X" and "did Y ship yet".
• Deep Research — runs dozens of searches iteratively, reads deeper, and returns a long structured report. Minutes rather than seconds. Use it when you would otherwise open twenty tabs.
• Labs — turns a prompt into an artifact: a spreadsheet, a small dashboard, a simple web app, a slide deck. Useful for "compare these fifteen vendors on these six axes" work.
• Spaces — a persistent collection with its own uploaded files and custom instructions. This is where the tool stops being a lookup and starts being a project workspace.
• Comet — a Chromium browser with the assistant built in: summarize the page, reason across open tabs, and run agentic tasks on your behalf. Free to download on all major platforms, with a Comet Plus tier for premium publisher content that Pro and Max already include.

Standout capabilities

• Inline citations on essentially every claim, with the source list always visible
• Focus filters that restrict retrieval to academic papers, or to a domain you name
• File upload (PDF, CSV, docs) so a question can span your files and the live web
• Model selection on paid tiers across the major frontier families
• The Sonar API, which exposes the same grounded, citation-carrying search to your own applications — the cheapest way to add "answers with sources" to a product without building a retrieval stack

Pricing shape

A capable free tier; Pro around $20/month for higher limits, model choice, and generous Deep Research use; a Max tier an order of magnitude above that for heavy users, bundling early features; discounted Education pricing; and per-seat Enterprise plans with the usual controls. The Sonar API bills separately per request and per token. See perplexity.ai/pricing for current numbers.

Best for

Anyone whose real job is finding out — analysts, journalists, engineers evaluating libraries, anyone writing a document where a wrong fact is expensive. It is also the fastest way to answer "has this changed since my model's training cutoff", which is a question you should ask far more often than feels natural.

Where it struggles

• Citation ≠ correctness. It will cheerfully cite a content-farm blog that copied a mistake. Check the source, not just the presence of a link.
• Long creative or agentic work. It is a research surface; it is not where you write the novel or refactor the service.
• Paywalled and login-walled sources are largely invisible to it, which quietly biases coverage toward whatever is free to crawl.
• Recency skew — fresh, heavily-SEO'd pages outrank the definitive older source more often than you'd like. Add "primary source" or a date range.

Getting more out of it

1. Ask for the shape of the answer: "compare these three on price, license, and maintenance cadence, as a table, with a source per cell."
2. Name your source preference — "prefer the vendor's own docs and filings over blog posts" measurably improves the citation quality.
3. Follow up with "what would contradict this?" — the second pass surfaces the dissenting sources the first one smoothed over.
4. Put standing context (your stack, your constraints, your audience) in a Space instead of retyping it every session.

Worth knowing

Treat Deep Research output as a well-organized reading list with a summary attached, not as a finished document. The value is that it read forty pages so you only have to read the six that matter — the six still need reading.