---
name: Perplexity Research
description: Use an answer engine like Perplexity for fast, cited research that speeds discovery without outsourcing judgment — follow the citations to the primary source, separate what's grounded from what's synthesized, and never ship a claim you didn't verify. For anyone using AI search to learn a domain or check a fact.
audience: everyone · developer · researcher
---

# Perplexity Research

## What this is

A method for getting real value out of a citation-first answer engine (Perplexity and its kind) as a research accelerator — entering an unfamiliar domain, gathering current information, checking a claim — while keeping the discipline that makes it trustworthy: the answer is a map to sources, not the destination. It is fast at finding; you are still responsible for verifying.

## What this is NOT

Not affiliated with Perplexity and not a claim about a specific model or feature — capabilities change, so confirm. Not a primary source, and not a substitute for reading one when it matters. A cited answer is a starting point that shows its work; it is not a guarantee the work is right.

## Method

1. **Ask a real question, with scope.** Specific, bounded queries get specific, checkable answers. "What changed in X between version A and B, with sources" beats "tell me about X."
2. **Follow the citation to the source.** The value of a cited answer is that you can check it — so check the ones that matter. A claim you're going to act on gets traced to the primary source, not accepted because a footnote exists.
3. **Separate grounded from synthesized.** Distinguish what a source actually says from the engine's paraphrase or inference across sources. The synthesis can be wrong even when every cited source is real — the join is where errors hide.
4. **Weigh the source, not just its presence.** A citation to a vendor blog, a forum post, and a standards body are not equal. Check the authority and the date; a confidently-cited stale source is still stale.
5. **Use it for discovery, confirm for decisions.** Perfect for mapping a space quickly, finding the right primary sources, and surfacing what you didn't know to ask. For anything you'll ship, decide, or quote — verify independently.
6. **Cross-check the surprising.** If an answer is surprising, load-bearing, or contradicts what you knew, get a second, independent source before believing it. One confident answer is not corroboration.
7. **Keep sensitive queries out.** Don't paste secrets, personal data, or confidential context into a third-party engine. If the question reveals something private, it doesn't go in the box.
8. **Cite what you carry forward.** When research feeds a doc, a decision, or code, keep the real sources with it — so the next reader can verify, and "we researched it" means something checkable.

## Quality bar

Queries are specific and scoped · load-bearing claims are traced to primary sources · grounded facts are separated from synthesized ones · sources are weighed by authority and date · discovery is fast but decisions are independently verified · surprising claims are cross-checked · sensitive data stays out · sources are carried forward with the conclusion.

## Guardrails & escalation

Regulated-domain questions (medical, legal, financial) from an answer engine are a starting question for a licensed professional, never an answer — route them. Anything destined for publication, a legal filing, or a product claim passes a human fact-check against primary sources; an invented or misread citation that ships is your error, not the engine's. When accuracy is critical and the sources disagree, escalate to a domain expert rather than picking the answer you like.

## References

- Catalogue: https://edwson.com/consumer-design-system.html · Contracts: https://edwson.com/cds/components.json · Agent brief: https://edwson.com/cds/AGENTS.md
- Related: the portfolio's "On using AI, honestly" method note and the CDS sources-&-citations entry (attach sources to the claim; label the ungrounded). Confirm Perplexity's current features and data handling against its own documentation.