What this covers
The “just ask ChatGPT to research X” approach produces something that looks polished and is half-fabricated — fake citations, blended sources, confident wrong numbers. A real research workflow uses ChatGPT for the three things it’s actually good at (outlining unfamiliar terrain, pointing at potential sources, synthesizing material YOU have read) and keeps you in the loop for the one thing it’s bad at (judging source quality). This guide is for analysts, researchers, and writers producing a briefing that has to survive scrutiny.
Who this is for
Anyone doing background research where the output will be read by someone who can call you out — an essay graded by a professor, a deal memo your MD will tear apart, a market-entry brief your CEO will quote. If the audience won’t check the sources, you don’t need a workflow; you need a paragraph.
When to reach for it
- Comparing 3-5 options where you need a structured trade-off table.
- Gathering background on an unfamiliar industry or technology before a meeting.
- Writing a short brief (1-3 pages) where every claim needs a source.
- Triaging which 10 papers to actually read out of 50 search results.
Before you start
- Define the deliverable shape first. “A 2-page brief with 8 cited claims and a recommendation” beats “research X.”
- Decide who the reader is and what they’d already know. The outline depth follows from that.
- Have a notes file open for source URLs and direct quotes. Don’t let evidence live only inside a chat.
- Turn on web search explicitly — don’t trust ChatGPT to know whether the topic needs current data.
Step by step
-
Ask for an outline of the topic at the right depth for your reader:
Outline the 5 main sub-topics of \{topic\} that a senior product manager needs to understand before a meeting with a vendor. For each, give one sentence on why it matters. -
For each section, ask web search for 3-5 potential sources, with publication dates:
For sub-topic 2, list 5 sources published in the last 18 months. Include title, author, publication, date, and a one-line summary. -
Open the top sources yourself. Read them. Paste the key paragraphs back into the chat — this is the step that prevents hallucinated citations.
-
Ask for synthesis grounded in the pasted text only:
Based ONLY on the paragraphs I pasted above (not your training data), write 4 bullets capturing the main agreements and 2 bullets capturing the main disagreements between these sources. -
End with a draft that cites the URLs you pasted, not citations the model generated for you. The model can rephrase your evidence; it cannot be trusted to manufacture it.
A prompt that produces honest output
You are helping me build a research brief on {topic}.
Constraints:
- If you don't know, say "I don't know — search needed."
- Never invent a citation. If you can't find a source, say so.
- Every claim must be either traceable to a source URL I'll provide,
or labeled "model opinion, not sourced."This dramatically reduces fabrication in my own usage; not zero, but noticeable.
Quality check
- Click every cited URL. Does the page exist? Does it actually say what the model claims?
- Cross-reference numbers against the primary source, not a summary article.
- Ask “what’s the counter-argument?” If the model can’t produce one, the brief is one-sided.
- Check publication dates. “Recent” in ChatGPT’s mind sometimes means 2023.
How to reuse this workflow
- Build a
research-template.mdwith your standard outline prompts and the anti-fabrication preamble. - Save successful research chats with descriptive names — they become a model for next time and an audit trail.
- For domains you research often (specific markets, regulators, technologies), keep a per-domain glossary in a Project for consistent terminology.
Recommended workflow
Outline → source-by-section (web search ON) → human reads top sources → paste key paragraphs → synthesis grounded in pasted text → cite the URLs you verified → final read with adversarial questions.
Common mistakes
- Skipping the source-verification step. Citations look real and are routinely wrong.
- Letting ChatGPT generate citations without you reading the underlying source — the model is fluent at fabricating plausible-looking references.
- Topic too broad to outline well. “AI in healthcare” needs to be narrowed to “AI scribing in primary care” before research starts.
- Mixing the model’s training-data knowledge with web-search results without distinguishing them. The brief ends up dated.
- Trusting tables it generates (“here are 5 vendors with prices”) without verifying each row. Half the rows are usually right; the bad ones are confidently wrong.
- Forgetting that any peer-reviewed paper takes precedence over a Medium post or LinkedIn carousel, even if the latter sounds more confident.
FAQ
- Should I use Perplexity or ChatGPT for research?: Perplexity is purpose-built for source-grounded research and produces better citations. ChatGPT is better for outlining and synthesis. Use both.
- What about Deep Research mode?: When available, it’s worth it for multi-source briefs. Treat the output as a first draft, not a final brief.
- How do I avoid getting cited fake papers?: Demand DOIs or arxiv IDs. Click through every one. If the model lists a paper without a working link, assume it’s invented.
- Can I skip reading the sources myself?: Only if you also accept being wrong in print. The model paraphrases, and paraphrases drift.