ChatGPT Deep Research Mode: The Professional’s Complete Guide

What ChatGPT Deep Research Mode Does

ChatGPT Deep Research mode is an agent that plans a research task, reads across many sources, and returns a structured report with citations you can check. Standard chat answers from what the model already knows. Deep research goes away for five to thirty minutes and comes back with something documented.

According to OpenAI’s Help Center documentation for deep research, a run moves through five stages:

1. You describe the outcome you need
2. You choose which sources it may use — the open web, files you upload, or connected apps
3. ChatGPT proposes a research plan, which you can review and edit before anything starts
4. You watch it work and can interrupt to refine the focus or change its sources
5. You receive a structured report with citations or source links

Why the wait is the feature

Most guides describe ChatGPT Deep Research mode as ChatGPT browsing faster. It is closer to the opposite. The agent reads a page, decides what it still does not know, and goes looking again — repeating that loop dozens of times. That iteration is what produces a report instead of a summary, and it is the same agentic pattern behind ChatGPT’s Agent Mode for multi-step tasks.

Finished runs open in a fullscreen report view with a table of contents, a list of every source used, and an activity history showing how the research progressed. You can download the result as Markdown, Word, or PDF.

Which Plan You Actually Need

Every paid tier includes ChatGPT Deep Research mode; the difference is how much of it you get. Free and Go accounts get limited runs, which is enough to test the workflow but not enough to build a habit around.

Plus costs $20 a month. Pro comes in two tiers — $100 for roughly five times Plus usage and $200 for twenty times — and both include deep research, so the choice there is headroom rather than capability. Business is $20 per user per month billed annually, or $25 monthly, starting at two users. If you are weighing the first upgrade, our breakdown of what ChatGPT Free and Plus actually include goes deeper on the rest of the feature gap.

One caveat worth knowing before you promise your team anything: availability depends on both your plan and your country.

Where to find it

There are three ways into ChatGPT Deep Research mode, and knowing all three saves the usual thirty seconds of hunting:

• Type /Deepresearch straight into the prompt box
• Pick Deep research from the tools menu, behind the + button
• Pick Deep research from the sidebar menu

The Research Commission Method

The Research Commission is a three-part brief that treats ChatGPT Deep Research mode as an analyst you are hiring rather than a search you are typing: state the Commission, set the Boundary, and specify the Blueprint. It maps directly onto the three things the tool actually asks you for, which is why it works.

Commission: name the decision

A topic produces a summary. A decision produces a recommendation. “Research the project management software market” is a topic. “Help me decide which of these three tools to shortlist for a 40-person operations team” is a commission, and the report writes itself differently as a result.

Boundary: control the sources

This is the control competitors rarely mention and the one that most improves output quality. From the prompt window, choose Sites then Manage sites to enter the domains you trust, as a comma-separated list. You then pick one of two modes: restrict the run to only those sites, or prioritise them while still allowing a full-web search.

Deep research can also pull from connected apps — document stores like Google Drive and SharePoint, and authenticated data sources including FactSet, PitchBook and Scholar Gateway. Note the guardrail: it only uses read actions from connected apps, never write actions.

Blueprint: specify the report

If you do not name the sections, you get whatever structure the model picks — and that is the single most common cause of a report you cannot paste into a deck. Number the sections. Cap their lengths. Say where you want a table.

COMMISSION
I need to decide [the specific decision] by [date]. The audience for this report is [role].

BOUNDARY
Use only: [named source types — official product sites, regulator publications, filed accounts, named review platforms].
Do not use: [vendor blog posts, undated listicles, press releases].
Where a claim cannot be verified from an allowed source, say "not verified" rather than estimating.

BLUEPRINT
Return the report in exactly these sections:
1. Recommendation (150 words, lead with the answer)
2. Comparison table across [criteria A], [criteria B], [criteria C]
3. Evidence for each option, one short paragraph each
4. What I still need to check before deciding
Cite every factual claim inline with the source URL.

That template borrows its bones from the Role, Goal, Context, Format structure behind the Logic System — the difference is that a research commission also has to govern where the facts come from, which ordinary prompts never need to do. Save your filled-in version somewhere you will find it again; if you would rather start from a ready-made set, our free AI work templates library has more briefs built this way.

Run a Competitor Brief in 25 Minutes

Here is the whole ChatGPT Deep Research mode workflow on a real task: a competitive brief for a leadership meeting. Roughly 13 of those 25 minutes are hands-on; the rest happens while you do something else entirely.

Draft it in a notepad first. Two minutes of thinking about which decision the report serves saves you a whole wasted run, and long pastes into the composer sometimes convert into a file attachment that the agent then treats differently from typed instructions.

COMMISSION
I am a product marketing manager preparing a competitive positioning brief for our leadership team, presenting Tuesday. Compare [Competitor A], [Competitor B] and [Competitor C] against us in the [your market] category.

BOUNDARY
Use only: each competitor's official website and pricing pages, their published documentation, filed financial statements where public, and named review platforms with dated reviews.
Do not use: vendor-sponsored comparison posts, SEO listicles, or any page without a visible date.
If pricing cannot be confirmed on an official page, write "not published" instead of estimating.

BLUEPRINT
Return exactly these sections:
1. Headline positioning for each competitor, one sentence each
2. Feature comparison table across [capability 1], [capability 2], [capability 3]
3. Published pricing and packaging, with the date each page was last updated
4. Two credible weaknesses per competitor, each tied to a dated source
5. Three positioning implications for us
Cite every claim inline with a full URL.

Open Sites then Manage sites and add the competitor domains plus any review platform you actually trust. For a competitive brief, prioritising those sites while allowing full-web search usually beats a hard restriction — you still want it to catch a recent news story you did not know about.

ChatGPT proposes a research plan, and may ask clarifying questions first. This is the most valuable 90 seconds in the whole process. Read the plan for the thing you did not ask for and the thing it missed, then edit it directly.

You can close the tab. If you are watching and see it wandering into the wrong territory, interrupt to refine the focus or adjust which sources it can reach — that is what the control is for. Otherwise leave it alone.

Download as Word or Markdown rather than copy-pasting from the chat window — the section structure survives the trip. From there, our guide to turning a long document into an executive summary covers the compression step for leadership audiences.

For the deeper competitor-research workflow specifically — scoring frameworks, what to track over time, how to structure the comparison — see our dedicated walkthrough on running competitor research with ChatGPT Deep Research.

Four More Jobs It Handles Well

The same three-part commission transfers to any task where the answer lives across many documents rather than inside one. What changes is the boundary and the blueprint, never the shape of the brief — which is why ChatGPT Deep Research mode rewards building one good template rather than improvising each time.

COMMISSION
I am recommending one of two vendors to our IT steering group and need to document the security position of each. The audience is a non-technical steering group.

BOUNDARY
Use only each vendor's official trust centre, security documentation, published certifications, and their current terms of service. Do not use analyst blog summaries.
Where a certification cannot be confirmed on the vendor's own site, mark it "unconfirmed".

BLUEPRINT
Return exactly these sections:
1. Risk table: certifications held, data centre regions, uptime commitment, training on customer data
2. Three potential deal-breakers per vendor, plain language
3. Questions to put to each vendor before signing
Cite every claim inline with the source URL and the date the page was last updated.

When the material is already sitting on your own drive rather than out on the web, the workflow differs — our guide on synthesising several documents into one brief covers that case properly.

What to Feed It Safely

ChatGPT Deep Research mode gets genuinely useful the moment you point it at your own material, which is exactly when the data question stops being theoretical. On Free, Go, Plus and Pro accounts, conversation content may be used to improve models, with an opt-out available in your settings. Business and Enterprise workspaces do not train on your data by default.

That difference matters more here than in ordinary chat, because a research commission tends to carry your actual strategic question inside it — which competitor you are worried about, which vendor you are about to drop.

Our full guide to using ChatGPT safely with work information covers retention settings and the workspace differences in more detail.

Verify Before You Send It Out

A citation proves the agent visited a page. It does not prove that page supports the sentence attached to it. That distinction is the single most important thing to understand about ChatGPT Deep Research mode, and skipping it is how a confident-looking brief embarrasses someone in a meeting.

Most strategy leads who use this well treat the citation list as an audit trail rather than a guarantee. Budget five minutes. You are not re-reading everything — you are checking the handful of claims that would actually change the decision if they turned out to be wrong.

Go back through the report you just produced and mark every factual claim as one of three things:

CONFIRMED — the cited source directly states this claim
INFERRED — you reasoned this from the source but the source does not state it
UNVERIFIED — you could not confirm this from an allowed source

List every INFERRED and UNVERIFIED claim separately at the end, with the sentence quoted, so I can check those first.
Do not rewrite the report. Only classify.

Anything landing in that list gets checked by opening the source yourself. If the underlying question of why AI states wrong things confidently is new to you, start with our explanation of why ChatGPT hallucinates and how to catch it, then our fact-checking workflow for AI research output.

When a Research Run Goes Wrong

Most ChatGPT Deep Research mode failures trace back to a vague commission rather than a broken tool. These are the patterns worth recognising quickly, so you stop a bad run early instead of waiting twenty minutes for a disappointment.

The interrupt control deserves a specific correction, because plenty of guides describe it wrongly. It exists to redirect a run — refining the focus or changing which sources it can reach. Treat it as steering, not as an emergency save button.

What Deep Research Still Cannot Do

This is the part most guides skip, and it is the part that decides whether you look competent using ChatGPT Deep Research mode or merely busy. The failure modes below are structural — no amount of better prompting removes them.

There is also a plain economics question. A run costs you 10 to 15 minutes of attention. For a question you could answer in three browser tabs, that is a bad trade, and reaching for deep research out of habit is the most common way people waste their allowance.

ChatGPT Deep Research mode has been available since early 2025, when OpenAI’s original deep research announcement launched it to Pro subscribers first. The controls have moved on considerably since — source scoping and editable plans in particular — but the underlying trade has not changed. You are exchanging your time for the machine’s, and the exchange only pays when the question genuinely spans more sources than you would open yourself.

Run ChatGPT Deep Research mode once this week on a real question you have been putting off. Write the commission properly, set a boundary, name your sections, and run the verification pass before the report goes anywhere near a colleague. The second one takes half as long as the first.