TL;DR
ChatGPT Deep Research is an agent that browses the live web for 5-30 minutes and returns a cited, multi-section brief. The output looks authoritative whether or not it is. This workflow keeps it honest: anchor context in a Project, force ChatGPT to challenge your question before it runs, review and edit the research plan, then spot-check real citations after. As of June 2026 Deep Research runs on a GPT-5.2-based model (the lightweight fallback uses o4-mini), and you get 25 runs/month on Plus (15 of them lightweight), 50 on Pro $100, and 250 on Pro $200.
What Deep Research actually is
Deep Research is a multi-step research agent inside ChatGPT. You give it a question; it plans a search strategy, issues sequential web queries, reads the pages and PDFs it finds, and synthesizes a structured report with inline citations. A run takes 5-30 minutes. It is not the same as a normal chat: a normal reply is instant and (often) un-sourced, while Deep Research goes away, reads hundreds of pages, and comes back with a brief.
Two things changed in early 2026 that this guide relies on:
- Editable research plan (since February 2026). Before the run starts, ChatGPT shows the plan — which angles it will take, what sources it intends to check, the output format. You can see it and edit it. This is the cheapest place to steer the whole run, so use it.
- Private-doc connectors (since March 2026). Deep Research can pull from SharePoint, OneDrive, and Dropbox alongside the public web, so a run can combine your internal documents with external evidence.
One thing has not changed: OpenAI itself states Deep Research “occasionally makes factual hallucinations or incorrect inferences.” The report carries citations, but it does not independently verify them. That is why the spot-check step below is non-skippable.
Who this is for
ChatGPT Plus, Go, Team/Business, or Pro users who want a defensible research brief — not a polished summary that falls apart when someone clicks a citation.
When to reach for it
When the question needs 30-90 minutes of web reading and a 5-15 page brief at the end — market sizing, competitive landscapes, regulatory scans, literature reviews. For a single quick fact, a normal chat or a web search is faster and cheaper against your quota.
Deep Research run limits (as of June 2026)
| Plan | Price/mo (USD) | Deep Research runs/mo | Notes |
|---|---|---|---|
| Free | $0 | 5 (lightweight only) | Lightweight = o4-mini, shorter briefs |
| Go | $8 | 25 (15 lightweight) | Same allowance as Plus |
| Plus | $20 | 25 (15 lightweight) | After the 10 full runs, it auto-drops to lightweight |
| Team / Business | per seat | 25 (15 lightweight) | Per-seat allowance |
| Pro | $100 | 50 | Same models as $200, 5x quota |
| Pro | $200 | 250 | 20x quota, full 1M-token in-app context |
Limits reset on a rolling 30-day window from your first run, not on the calendar month. When your full-run allowance is gone, queries silently fall back to the lightweight model — so a “Deep Research” answer that comes back unusually fast late in the month is probably the lightweight version.
Step by step
-
Create a Project to anchor context. ChatGPT sidebar →
+ New project. Name it after the deliverable, e.g.2026 SaaS APAC pricing research. Project header →...→Instructions, paste:Your role: senior research analyst, specializing in [your domain]. Research goal (the core question this project answers): [one line, e.g. "Median and distribution of annual contract value for B2B SaaS in APAC, 2026"] Audience: [e.g. "me + 2 investors, all product-literate but not APAC experts"] Deliverable: [e.g. "8-section brief, 600 words per section, each section cites sources"] Before every answer: 1. Restate my sub-question (one sentence max) 2. List which project sources you used this turn STOP and ask me when: - Information is insufficient to conclude (do not invent numbers or quotes) - My new question conflicts with the project goal -
Upload seed material. Project right rail →
Files→Add files:- Existing decks, old briefs, any prior research (PDF / MD / DOCX)
- 3-5 authoritative URLs you already trust, saved as
seed_sources.md(one URL per line plus one sentence on why it is authoritative) - A
glossary.mddefining the domain terms you want ChatGPT to use, so it does not fall back on generic explanations
If your evidence lives in SharePoint, OneDrive, or Dropbox, connect those instead of re-uploading — the March 2026 connectors let a run read them directly.
-
Refine the question — make ChatGPT challenge you first. Open the first chat in the project:
Research goal is in instructions. Based on seed_sources and glossary, before I run Deep Research, do exactly these 3 things: 1. Decompose the research goal into 5-8 sub-questions, each independently researchable, 20 words max 2. For each sub-question, list: - Which seed source partially answers it - What source type is still needed (industry report / 10-K / academic / primary interview) 3. Challenge me: what is ambiguous in the goal, what angles am I likely missing, what hidden assumptions could invalidate the conclusion Do NOT answer the research goal itself. Only output the 3 blocks above.Manually drop fuzzy sub-questions, add the angles you missed, then proceed.
-
Run Deep Research, and edit the plan before it starts. Same project, new chat. In the model picker choose a Thinking-class model and toggle
Deep Researchon (the exact label tracks the current ChatGPT build — as of June 2026 the underlying model is GPT-5.2-based). Send:Research goal (from project instructions): [paste the 1-line goal] Refined sub-questions: 1. [sub-question 1] 2. [sub-question 2] ... Deliverable spec: - 8 sections, each section title = one sub-question - 600 words per section max - 3-6 inline citations per section - End with a `## References` block: link + publication date + source type (report / 10-K / blog / primary) - Every number must have a source; if no primary source exists, explicitly write "not found" - Mark each finding as: "established" (2+ independent sources agree) / "contested" (sources conflict) / "uncertain" (only 1 low-authority source) Time scope: only 2024+ sources as primary evidence.ChatGPT will show a research plan before it runs. Read it. If it is heading for the wrong sources, the wrong time window, or skipping a sub-question, edit the plan now — this is the single cheapest steering point in the whole run. Then start it. The run takes 5-30 minutes; do not send other messages in this chat while it works (a new message cancels the run).
-
Mark weak sections, then re-research. When the output lands, scan it and mark a section “weak” if:
- Citations are all blogs / Medium / Reddit, no primary data
- Numbers are given without primary-source links (hedge words like “approximately”, “estimated”)
- The sub-question is dodged or answered too generically
Pick the 2 weakest, open a new chat (same project, Deep Research still on):
Below is section [N] from the prior Deep Research run: [paste the section] Redo this section. Rules: 1. Replace all citations — only use these source types: industry reports (Gartner / Forrester / IDC / commissioned research) / public-company filings / primary data (expert interviews, community surveys, official API data) 2. No Medium / Reddit / personal blogs as primary evidence 3. Every number needs a clickable original source link (not "as reported by X") -
Citation spot-check — the one step you cannot skip. Because Deep Research synthesizes from web content without verifying it, you verify. Pick 5 inline citations at random, and for each:
- Click the link, confirm the page exists (not a dead link or paywalled)
- Use Find (Cmd/Ctrl+F) on the original page for the cited number or phrase, and confirm it actually appears
- Flag the section “weak” if: the original says the opposite / it is a secondhand citation (the source is quoting an earlier source) / the source is more than 2 years stale
For every flagged section, redo it via step 5.
-
Executive summary — with explicit confidence labels:
Based on all 8 sections plus the redone sections, write a 1-page executive summary: - Length 250 words max - Structure: - "Established" paragraph (multi-source corroborated, 2+ conclusions) - "Contested" paragraph (sources conflict — name the conflict, do not pick a side) - "Uncertain" paragraph (missing data or single source — list what is still needed) - "Next steps" paragraph (3 actionable items max) Tag each conclusion in brackets with confidence: [established / contested / uncertain]. No "significant" or "substantial" without a magnitude. -
Archive and reuse. Bundle the entire project (instructions, seed sources, refined sub-questions, Deep Research output, citation audit, executive summary):
chatgpt_research_2026_05_21_saas_pricing/ ├── instructions.md ├── seed_sources.md ├── refined_subquestions.md ├── deep_research_v1.md ├── deep_research_v2_redone_sections.md ├── citation_audit.md └── executive_summary.mdFor the next similar project: duplicate the Project, change the “Research goal” in the instructions, swap the seed sources. Steps 1-3 then collapse into about 5 minutes.
Common mistakes
- Treating the first Deep Research draft as final — it is a draft until the citations are checked.
- Skipping the editable plan and letting it run on the wrong sources or time window.
- Not using Projects, so context is lost across runs and every run starts cold.
- Skipping the “challenge my question” step, so the brief answers a vague question precisely.
- Not labeling “contested vs settled”, so a single conflicting source quietly becomes a “fact”.
- Burning full-run quota on trivial lookups, then hitting lightweight-only for the real briefs.
FAQ
- Deep Research vs a normal ChatGPT chat? Deep Research browses the live web for 5-30 minutes and writes a long, cited brief. A normal chat is instant but draws mostly on the model’s training plus whatever single search it runs, with weaker sourcing.
- How many Deep Research runs do I get? As of June 2026: Free 5/month (lightweight only), Go and Plus 25/month (15 lightweight), Team/Business 25/month per seat, Pro $100 50/month, Pro $200 250/month. Quotas reset on a rolling 30-day window from your first run.
- What is the “lightweight” version? When your full-run allowance is exhausted, runs fall back to a lighter o4-mini-based model that produces shorter, shallower briefs. It is fine for quick scans, not for a brief you will defend.
- Can it read my private files? Yes — upload them to the Project, or use the SharePoint / OneDrive / Dropbox connectors added in March 2026 so a run can combine your internal docs with public sources.
- Why does my run come back in under a minute sometimes? That is usually the lightweight fallback (allowance spent) or a question that needed little browsing. A real multi-source run takes several minutes.
- Do I still have to fact-check it? Yes. OpenAI states Deep Research occasionally hallucinates or makes incorrect inferences and does not verify the sources it cites. The spot-check in step 6 is the floor, not optional polish.
Related
External references: OpenAI — Introducing deep research · ChatGPT pricing and plans