How to Write a Perplexity Research Brief with Source Policy
A Perplexity run without a source policy is a citation lottery. The model will mix your paste, a remembered paper, and a URL that never existed. You send the brief. A reviewer clicks the first link. It 404s. The job is a brief that echoes the policy you set, quotes only sources you allowed, and prints a NO_DATA list instead of a fake bibliography.
The matching generator is the Perplexity Research Brief with Source Policy prompt. Browse related cards in the PromptDig library (Browse more prompts). When a filled run survives a source check, share the version you actually use (Share a prompt).
Echo the policy before you write findings
Paste the question, the sources you may use, and the pastes or files you actually have. Source policy is the rule set: primary documents only, no blogs, no paywalled PDFs you did not paste, no "generally known" facts. Output shape is the brief you need: one pager, table, or memo with a findings list. Off-limits is anything you will not accept: invented DOIs, unnamed analysts, competitor quotes you did not paste.
The first block the generator should return is a source policy echo. If you said only the three URLs in Sources, the echo must say that. If it quietly expands to plus well-known industry reports, stop. That expansion is how fake citations get in.
Findings may use Pastes and Sources only. A sentence that needs a fourth URL is a gap, not a chance to invent a McKinsey PDF. Dates, sample sizes, and quotes need a pointer back to a named source in Inputs. "Studies show" is not a pointer.
Refuse invented citations, DOIs, and quotes
Citations: only items that appear in Sources or Pastes. Do not complete a half-remembered APA line. Do not mint a DOI because the title sounds real. If a paste has a title and no DOI, the brief keeps the title and marks DOI as NOT IN INPUTS.
Quotes: only text that is in Pastes, attributed to the speaker or document named there. Do not polish a paraphrase into quotation marks. Do not attribute a finding to Perplexity's index.
Numbers: only figures that are in Pastes or in a source you named and pasted. Do not annualize, convert currencies, or update a 2022 figure to 2026. If the paste says about thirty, write about thirty. Do not print 31.4%.
If the model names a tool version, lock the version you named. If you did not name a Perplexity mode or date, write unknown. Do not pretend the run used a specific index snapshot.
Method notes, gaps, and a brief that can survive a click-through
Method notes should say what you searched, what you pasted, and what you refused. That is the audit trail. If you only pasted two PDFs, the method is two PDFs, not a comprehensive scan.
Gaps belong on the page: missing dates, missing primary URLs, missing sample sizes, missing jurisdiction, missing whether a source is paywalled. Print them before anyone forwards the brief. A one-pager with five gaps is more useful than a two-pager with five invented links.
Banned words belong on the card: best, guaranteed, #1, invented percents, celebrity. The compliance pass should quote leaks and cut them. If the question is legal, clinical, insurance, HR, veterinary, or education-plan adjacent, add a one-line not-advice banner and de-identify emails, phones, and record IDs in the paste.
A worked pass: Friday deadline, thin paste
Say the question is a vendor comparison, Sources lists two product pages, Pastes has a deadline of Friday, an unapproved budget, and no customer quotes. The brief can restate those two pages and the Friday deadline. It cannot invent a Gartner quadrant, a list price, or a champion quote. The NO_DATA list should include budget figures, quotes, and any third vendor you did not paste.
If a later run gets a third page, paste it and rerun. Do not edit invented citations from a previous run. The source policy echo should change first. Then the findings.
When you later get a DOI from the publisher page you actually opened, add it to Pastes. Do not let the model fill in Crossref. The point of this card is a brief a skeptic can click through, not a denser footnote block.
A second failure mode is the model "helpfully" merging two papers into one citation. If Pastes has Paper A without a year and Paper B with a year, do not produce Paper A (2021). Split them. Mark the missing year. Reviewers notice fused citations faster than they notice a missing comma.
If Off-limits named blogs and the model still cites a Substack, that is a policy miss. Put the URL in the refuse list. Do not argue that the author is an expert. The policy is the policy.
Fill the card, then run
Name the question, the sources you allow, the files you pasted, and the shape you need. Leave Off-limits explicit.
Question: [Question]
Sources I may use: [Sources]
Pastes / files: [Pastes]
Source policy: [Policy]
Output shape: [Shape]
Off-limits: [OffLimits]
A research brief is a chain of custody, not a vibe. When a run keeps a DOI blank because you never pasted one, share that filled card. That is the adapter other analysts actually want.