How to Build a Cited Research Report with Perplexity Comet
Most AI research output falls apart at the citation stage. Numbers appear without sources, or sources do not say what the summary claims. Perplexity Comet can produce a properly sourced document, but only if you run a disciplined pipeline.
Short answer: collect primary sources as open tabs, have the sidebar extract claims with URLs into a table, reconcile disagreements yourself, then draft from that table. Never draft from a prose summary.
Why start from tabs instead of a search query?
A search query returns whatever ranks. Open tabs let you curate: you pick the regulator’s PDF, the vendor’s pricing page, and two independent analyses. Comet’s assistant reads exactly what you opened, which kills a whole class of sourcing error.
Aim for six to twelve primary sources. Fewer and you cannot cross-check; more and the context gets diluted.
How do you extract claims with their sources?
- Open your curated tabs in one window.
- Ask the sidebar: “From these tabs only, list every quantitative claim with the exact URL and the sentence that supports it.”
- Require a table format: claim, value, source URL, supporting quote.
- Export that table to your notes before doing anything else.
- Keep the tabs open while drafting so you can re-query individual sources.
The phrase “from these tabs only” matters. Without it, the assistant quietly brings in outside knowledge and your source trail breaks.
How do you handle sources that disagree?
This is the actual value of the workflow. Once you have the claim table, sort by topic and look for conflicting values on the same metric.
| Conflict type | Resolution |
|---|---|
| Different dates | Use the most recent, note the older figure |
| Different definitions | State the definition explicitly |
| Vendor vs third party | Prefer the independent number |
Flag unresolved conflicts in the report rather than picking one silently. Readers of research trust a document more when it names its uncertainty.
What does the drafting step look like?
Draft from the table, not from chat history. Each paragraph should map to one or two rows. If a sentence needs a fact that is not in the table, go get the source or cut the sentence.
Ask Comet to check the final draft as a separate step: paste the report and ask it to flag any claim without a citation. That catches sentence three of a paragraph where citation two quietly covers four claims.
Where does this workflow not pay off?
It is slow. Curating ten sources and building a claim table takes two to four hours for a mid-size report. For a quick internal memo, that is absurd overhead.
Use it for anything public, legal-adjacent, or financial — documents where a wrong number has consequences. For a Slack update, just ask the assistant and move on.
Quick Summary
- Curate six to twelve primary sources as tabs before asking anything.
- Extract claims into a table with URL and supporting quote, from those tabs only.
- Reconcile conflicting values explicitly instead of silently picking one.
- Draft from the claim table so every sentence maps to a source row.
- Worth the overhead for public or financial reports, not for quick memos.
