Using Deep Research (OpenAI) for a Rapid Literature Scan
A step-by-step workflow for using OpenAI's Deep Research to produce a cited landscape overview in 20–30 minutes, followed by a verification pass to catch hallucinations before you build on the output.
What you’ll accomplish
By the end of this tutorial you will have:
- A 1,500–3,000 word overview of a research area with citations, generated in under 30 minutes
- A verification workflow to check the output before building on it
- A clear understanding of when to hand off to Elicit for systematic work
Prerequisites: ChatGPT Plus or Pro subscription ($20–200/mo). Access to your institution’s journal database (for verification).
Step 1 — Frame your question precisely (5 minutes)
Deep Research performs best on questions with a defined scope. Before you start, write out:
- The topic — what field, intervention, or phenomenon you’re examining
- The angle — what specific aspect you need to understand (mechanisms, clinical outcomes, current tools, open debates)
- The depth — are you mapping the whole field, or drilling into one specific sub-question?
Good prompts for Deep Research:
- “Give me an overview of AI methods for predicting antibiotic resistance from bacterial genome sequences, focusing on what approaches have shown clinical validation and what limitations remain.”
- “Survey the current state of neuroimaging biomarkers for Alzheimer’s disease progression — what imaging modalities are being used, which have the strongest predictive power, and what are the main reproducibility concerns?”
- “What are the main AI tools for crystal structure prediction in battery materials, how do they work, and what does the experimental validation literature show about their accuracy?”
Prompts that give poor results:
- “Tell me about machine learning in medicine” — too broad
- “What are the latest papers on CRISPR?” — too open-ended; Deep Research doesn’t know what “latest” means to you
- Questions where the answer is in one specific paper — regular ChatGPT or Elicit is better
Write your final prompt before opening Deep Research.
Step 2 — Run Deep Research (20–30 minutes)
- Open chatgpt.com and start a new conversation
- At the bottom of the input box, click the Tools or Deep Research icon (it looks like a document with a magnifying glass; the exact UI varies by version)
- Paste your prepared prompt and submit
- Wait — Deep Research will show a live “thinking” and “searching” pane as it works. A typical run takes 10–30 minutes. You can close the tab and come back; the result is saved.
While waiting: read one background paper you already know on the topic. This primes your fact-checking in the next step.
Step 3 — First read: structure and scope (5 minutes)
When the report arrives, do a fast first read focused on structure, not accuracy:
- Does it cover the sub-areas you care about?
- Are there obvious major topics it missed?
- Does the organization make sense for your purpose?
If the scope is wrong, iterate the prompt before investing time in verification. Common adjustments:
- “Focus more on [X aspect] and less on [Y aspect]”
- “Add a section specifically on [missing topic]”
- “Rewrite this with more emphasis on methodological limitations”
Each follow-up runs another research pass (another 10–20 minutes).
Step 4 — Verification pass (15 minutes)
This step is non-optional. Deep Research produces plausible, fluent text with real-looking citations that are sometimes wrong. A 15-minute verification pass catches the most consequential errors before you build on the output.
What to check:
High-value claims — pick 5–7 specific factual claims that you would cite or build arguments on. For each:
- Click the inline citation link
- Read the source page to confirm the claim
- If the source doesn’t support the claim, flag it with a comment
Key statistics — any numbers (percentages, effect sizes, sample sizes, dates) should be verified against the source, not just checked for plausibility.
Paper citations — Deep Research sometimes cites papers that exist but don’t say what it claims they say. For papers you intend to cite in your own work, retrieve the abstract or full text and verify.
A realistic expectation: In a 2,000-word Deep Research report, expect 2–5 claims that are imprecise or unsupported, and 1–2 citations that misrepresent the source. This is not a reason to discard the output — it is a reason to verify before using it.
Step 5 — Export and build your reference list
Once you’ve verified the key claims:
- Copy the report text into a document
- For each verified citation, retrieve the actual paper (DOI, PubMed ID, or Google Scholar)
- Add verified papers to Zotero for proper citation management
At this point you have:
- A verified landscape overview
- A seed reference list of ~20–40 papers
Do not cite Deep Research directly in academic work. Cite the primary sources it pointed you to.
When to hand off to Elicit
Deep Research gives you a landscape; it does not give you a systematic review. After your Deep Research pass, use Elicit when you need to:
- Extract structured data from many papers (methods, sample sizes, outcomes) in a tabular format
- Screen against inclusion/exclusion criteria systematically
- Ensure coverage — Elicit queries Semantic Scholar’s database of 200M+ papers, not just what appears on the web
A practical handoff: take the key sub-topics and terminology you learned from Deep Research, run those as structured queries in Elicit, and build your systematic extraction from the more comprehensive Elicit results.
Common mistakes to avoid
Taking the report structure as authoritative. Deep Research organizes its output in a plausible way, but the structure reflects what it found and how it grouped it — not necessarily the most important divisions in the field. Your domain knowledge should dictate the final structure.
Citing before verifying. Even one unverified citation in a grant application or manuscript can create serious credibility problems if a reviewer spots it. Verify first.
Using Deep Research for clinical or safety-critical claims. For anything that affects patient care, regulatory submissions, or safety-relevant decisions, use primary literature accessed directly — not AI-synthesized summaries.
Skipping the follow-up literature search. Deep Research covers what is publicly accessible on the web. Paywalled papers, recent preprints, and specialized conference proceedings may be missing. Always follow up with a direct database search (PubMed, Semantic Scholar, arXiv) for completeness.
Time summary
| Step | Time |
|---|---|
| Frame your question | 5 min |
| Run Deep Research | 20–30 min (automated) |
| First read and scope check | 5 min |
| Verification pass | 15 min |
| Export and Zotero import | 10 min |
| Total active time | ~35 minutes |