How to Prompt ChatGPT Deep Research

TL;DR: ChatGPT Deep Research is OpenAI's report-writing agent, distinct from a normal chat turn or the quick Search tool. It proposes an editable research plan, may ask clarifying questions, runs for roughly 5 to 30 minutes, and returns a cited, exportable report. Write your prompt as a research brief, not a question, and open every citation that matters before you trust it. 28 copy-paste briefs are below.

What Is ChatGPT Deep Research, and What Does It Do Today?

OpenAI's own help center titles this feature "Deep research in ChatGPT," and its plain-language definition is worth quoting exactly: "Deep research helps you accomplish complex online tasks by reasoning, researching, and synthesizing information into a documented report" (help.openai.com, accessed September 2, 2026). It works from three defaults: the public web, files you upload, and connected apps you have enabled. You can narrow it to specific sites.

The feature launched under a different technical story than it carries now. OpenAI's original release post, dated February 2, 2025, opens with: "Powered by a version of the upcoming OpenAI o3 model that’s optimized for web browsing and data analysis, it leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet" (openai.com, accessed September 2, 2026). The current help center article drops the model name entirely: "Deep research is powered by the latest models for fast, accurate results. If you prefer, you can choose a legacy model for a deep research task." That shift matters for anyone whose prompts reference a specific model by name. The branding has already moved on once, and the doc itself now hedges against needing to move it again.

Every plan is not equal here. OpenAI's live pricing table marks Deep Research "Limited" on the Free and Go plans and unrestricted on Plus and Pro (verified against openai.com/chatgpt/pricing, September 2, 2026, read from the page's own accessible labels rather than a summary). None of the current pages attach a number to "Limited" or to the Plus and Pro allowance. An April 2025 update to the original release post did publish numbers at the time: Plus, Team, Enterprise, and Edu users getting a fixed monthly allowance, Pro users a larger one, Free users a small one, with overflow falling back to a lighter model. That update is now roughly a year and a half old, though, and the current help article replaces any fixed figure with "usage varies by plan" and an in-product counter. Treat any specific number you read elsewhere as dated unless it cites a page you can check today.

How Is Deep Research Different From a Normal Turn, Search, or Agent Mode?

ChatGPT now ships several tools that sound similar and behave differently, and OpenAI draws the first line itself: "Use search for quick facts, and use deep research for depth and thoroughness. Search quickly pulls recent web information and returns a short summary with links. Deep research takes more time to read and analyze many sources, then produces a detailed, documented report." A second surface, Agent mode, overlaps in ability but not in identity. A July 17, 2025 update to the same release post states that "original deep research functionality remains available via the “deep research” option in the tools menu" even after Agent mode shipped its own, browser-driven version of similar work.

That table is also the fastest way to catch a wrong prompt. If your question needs an answer before you finish reading this sentence, it belongs in Search. If it needs a person clicking through a real interface, that is Agent mode's job, not Deep Research's.

Why Should a Deep Research Prompt Read Like a Brief, Not a Question?

Because that is what OpenAI's own guidance says a good one looks like, almost word for word: "A good deep research prompt clearly describes the question, desired outcome, and any relevant constraints." A one-line question gives the model exactly one of those three things. Everything Deep Research does well (choosing a source class, holding a recency window, knowing what to leave out, shaping the final report) depends on constraints you supply up front, because there is no back-and-forth once the plan is running.

Here is the reusable shape:

Research question: [WHAT YOU NEED ESTABLISHED, as one specific question]
Decision this informs: [WHAT YOU WILL DO WITH THE ANSWER]
Preferred sources: [SOURCE CLASS — regulator filings, vendor documentation,
  peer-reviewed studies, company financial statements, primary reporting]
Recency: [WINDOW — "published in the last 12 months" or "current as of today"]
Exclude: [WHAT TO IGNORE — forums, SEO round-ups, a named competitor's own
  marketing copy, anything before a stated date]
Output shape: [TABLE WITH N ROWS / RANKED LIST / ONE PARAGRAPH PER SOURCE],
  plus a closing section listing anything you could not confirm

Fill in as many lines as the task deserves and drop the rest as a single paragraph. The point is not the format; it is that a decision, a source class, and an exclusion list are things a one-line question never carries, and a research plan that runs unattended for half an hour cannot ask you for them mid-task the way a chat model can.

Does Deep Research Ask Clarifying Questions, and Can You Edit the Plan?

Yes to both, and this is where the brief format pays off twice. OpenAI's help center: "Before research begins, ChatGPT may ask clarifying questions to confirm your goals. You can review and edit its proposed research plan so the report stays aligned with what you want to accomplish." A well-built brief answers most of those clarifying questions before they get asked, but the plan-review step is still worth reading closely: it is the last point where you can add an excluded source or a missed constraint before the task commits to 20 minutes of unattended work.

The same help page documents a second intervention point most users never use: "You can follow progress as it runs and interrupt at any time to refine the focus, including adjusting which sources it can access." If the plan review is your last chance to fix a bad brief, the interrupt is your first chance to fix a research run that has already started drifting, for instance if the sources it is reading turn out to be lower quality than you specified.

If you routinely rewrite the same clarifying answers, teaching ChatGPT to ask better questions up front is worth reading alongside this one, since the discipline transfers even though Deep Research's clarifying step behaves differently from a normal chat turn's.

What Does Deep Research Actually Browse, and What Doesn't It Touch?

Three default sources, per OpenAI's help center, listed as its own three bullet points: the public web, files you upload, and connected apps. Connected apps can include document stores such as Google Drive or SharePoint and authenticated industry data services, but availability depends on your plan, region, workspace settings, and the app's own permissions. An app appearing in ChatGPT elsewhere does not mean Deep Research can use it.

The access is deliberately one-directional. OpenAI states it without qualification: "Deep research uses available read actions from connected apps. It does not use app write actions as part of research." Nothing gets created, sent, or changed in a connected account as a side effect of a research task. It reads what it is permitted to read and stops there.

You can also narrow the web side specifically. Through Sites → Manage sites in the prompt window, you can restrict research to a list of domains, or choose a softer option that prioritizes named sites while still allowing the broader web. Put that restriction in your brief's "preferred sources" line if precision matters more than coverage, and expect a smaller but more targeted set of citations back.

One country-level caveat worth stating plainly, since OpenAI's own FAQ does: "Does deep research work in all countries? No. Availability depends on your plan and your country or territory." If a colleague reports different behavior than yours, plan and region are the first two things to check, not your prompt.

How Long Does It Run, and What Do You Get at the End?

OpenAI's original release post gives the only duration figure that exists on a primary source, and it is a range, not a promise: "Deep research may take anywhere from 5 to 30 minutes to complete its work, taking the time needed to dive deep into the web." No later document tightens that range or replaces it with something more specific, so treat 5 to 30 minutes as the honest answer rather than a number this post invented to look more precise than OpenAI's own docs.

What arrives at the end is more than a chat reply. The help center describes "a fullscreen report view" with three parts: "A table of contents for navigating long reports," "A sources used section for reference checking," and "An activity history showing how the research progressed." All three are worth reading before the prose. The activity history in particular shows you which searches actually ran, which is a faster way to spot a research task that drifted off-brief than reading the whole report looking for it.

Exportability is confirmed, not assumed: "You can download completed reports for reuse or sharing in multiple formats, including Markdown, Word, and PDF." That makes a Deep Research report a genuine work product rather than something you have to re-copy out of a chat window, which is part of why treating the prompt as a brief, something a document deserves, is the right mental model rather than treating it as a chat message that happens to run longer.

Is a Deep Research Citation the Same as a Verified Fact?

No, and this is the single most important habit in this post. A citation tells you a page was read during the research run. It does not tell you the sentence sitting next to that citation is an accurate rendering of what the page said. OpenAI's own limitations language, from the original release post, is honest about the underlying risk even though it gives no number: research "can sometimes hallucinate facts in responses or make incorrect inferences, though at a notably lower rate than existing ChatGPT models, according to internal evaluations." That is a comparison to OpenAI's other models, not a published accuracy rate for Deep Research itself, and no later document supplies one.

The practical fix costs you almost nothing given what the report view already provides: open the sources-used panel, and for any claim the report's output actually depends on, click through to the source and read the sentence yourself. This is the same discipline that keeps a literature review from losing its citations and the same one any AI research-and-citation workflow needs: a documented report is not a fact-checked one, no matter how many footnotes it carries.

28 Copy-Paste Deep Research Briefs

Paste one of these into the Deep Research composer, adjust the bracketed pieces, and expect a clarifying question or two before the plan appears. Each one is written as a brief: a decision, a source class, a window, and a shape, not a bare question.

Competitive and vendor research

1. Research question: What are [COMPETITOR]'s current published pricing tiers
   and their monthly and annual prices? Decision: whether to reposition our
   own pricing page. Sources: the vendor's own pricing and terms pages only.
   Recency: current as of today. Exclude: third-party pricing round-ups.
   Output: one table, one row per tier, quote the figure and note if a tier
   is not published rather than estimating it.

2. Research question: What has [COMPANY] announced publicly in the last 90
   days? Decision: whether it changes our competitive positioning this
   quarter. Sources: the company's own newsroom, blog, and regulatory
   filings. Recency: last 90 days only. Exclude: press aggregation sites
   that reword the original announcement. Output: chronological list, one
   line per announcement, each with its date and source URL.

3. Research question: Which vendors does [ANALYST FIRM] name in its most
   recent report on [CATEGORY], and on what date was that report published?
   Decision: whether to brief the sales team on a new entrant. Sources: the
   report itself if reachable, otherwise its own summary page. Exclude:
   secondhand coverage of the report. Output: a list of named vendors with
   one line of context each, and the report's publication date up top.

4. Research question: What does [COMPANY]'s own documentation say about
   [SPECIFIC FEATURE], and has that description changed in the last year?
   Decision: whether a comparison page we maintain is still accurate.
   Sources: the vendor's current docs, plus an archived version if you can
   reach one. Output: current wording, prior wording if found, and the date
   each was current.

5. Research question: What do verified user reviews on [REVIEW PLATFORM]
   say about [COMPETITOR]'s support quality in the last six months? Decision:
   whether "poor support" is a safe claim to make in our own marketing.
   Sources: the review platform directly, not summaries of it. Exclude:
   reviews older than six months. Output: a short list of representative
   quotes with dates, and a one-line summary of the overall pattern.

Regulatory and compliance

6. Research question: What does [REGULATION] require of a company in our
   position, and from what date does each requirement take effect? Decision:
   what our compliance checklist needs to add this quarter. Sources: the
   official statute or regulation text and the regulator's own guidance.
   Exclude: law-firm marketing summaries. Output: one row per requirement,
   with its effective date and the section number it comes from.

7. Research question: Has [REGULATOR] published any guidance on [TOPIC] in
   the past 12 months? Decision: whether our current process needs review.
   Sources: the regulator's own site only. Recency: last 12 months. Output:
   a list with title, date, and URL for each document found; if nothing was
   published, say so rather than substituting older material.

8. Research question: How does [JURISDICTION A] define [TERM] compared with
   [JURISDICTION B]? Decision: whether one contract template works for both
   markets. Sources: each jurisdiction's own statutory text. Output: both
   definitions quoted verbatim, with an explicit note anywhere the two are
   not directly comparable rather than a forced reconciliation.

9. Research question: What is the current status of [PENDING RULE OR BILL],
   and when is its next scheduled action? Decision: whether to delay a
   related product decision. Sources: the legislature's or regulator's own
   tracker. Output: current status, next date, and a one-line history of
   the last three actions taken on it.

Academic and literature synthesis

10. Research question: What do peer-reviewed human trials say about
    [INTERVENTION] and [OUTCOME] published since [YEAR]? Decision: whether
    to cite this in a report going to a non-specialist audience. Sources:
    peer-reviewed journals only; label any preprint clearly if included.
    Output: one paragraph per study — sample size, method, effect size,
    journal — plus a closing list of anything you could not verify.

11. Research question: Which papers cite [PAPER] and explicitly disagree
    with its conclusion? Decision: whether our internal summary of the
    original paper needs a caveat added. Output: one line per disagreeing
    paper naming the specific point of disagreement, not just that one
    exists.

12. Research question: What is the current edition or version of [STANDARD
    OR SPECIFICATION], and what changed from the prior one? Decision:
    whether our internal documentation cites a superseded version. Sources:
    the standards body's own publication page. Output: current version,
    prior version, and a short list of what changed between them.

13. Research question: What do the last three years of review articles say
    is still unresolved about [RESEARCH QUESTION]? Decision: scoping a new
    internal research effort. Sources: review articles in peer-reviewed
    journals. Output: a short list of open questions, each attributed to
    the review that raised it, with publication year.

Purchase and personal decisions

14. Research question: What do [PRODUCT A] and [PRODUCT B]'s own
    specification pages say about [THREE SPECIFIC ATTRIBUTES]? Decision:
    which one to buy. Sources: manufacturer specification pages only.
    Output: one table, one row per attribute; where a manufacturer does not
    publish a figure, write "not published."

15. Research question: What do the last six months of verified owner reports
    say about [KNOWN ISSUE] with [PRODUCT]? Decision: whether the issue is
    common enough to avoid the product. Sources: manufacturer acknowledgment
    if any, plus verified owner forums. Output: separate the vendor's own
    statement from user reports, and do not treat forum consensus as a
    confirmed defect rate.

16. Research question: What is [PRODUCT]'s published warranty and return
    policy in [COUNTRY]? Decision: whether to buy from this retailer or a
    different one. Sources: the manufacturer's or retailer's own policy
    page. Output: the policy quoted directly, with the date you retrieved
    it, since these change without notice.

17. Research question: What neighborhoods in [CITY] have seen the largest
    change in [METRIC — school ratings, commute time, listed home prices]
    over the last three years? Decision: where to focus a house search.
    Sources: municipal or official data portals. Output: a ranked list with
    the metric's value at the start and end of the window.

Industry trend and market sizing

18. Research question: What is the current estimated market size for
    [INDUSTRY OR CATEGORY], and which research firms have published a
    figure in the last two years? Decision: whether to include a market-
    size claim in an investor update. Sources: named analyst firms and
    their own published reports. Output: each figure with its source, year,
    and stated methodology if published; do not average conflicting figures
    into a new one.

19. Research question: Which three companies have raised the largest
    disclosed funding rounds in [SECTOR] in the last two quarters? Decision:
    scoping a competitive landscape slide. Sources: the companies' own
    announcements or primary funding databases. Output: company, amount,
    lead investor, announcement date, one line per company.

20. Research question: What do the last four quarters of [INDUSTRY]
    earnings calls from public companies say about [SPECIFIC TREND]?
    Decision: whether the trend is accelerating or fading. Sources: the
    companies' own transcripts or filings. Output: one line per company per
    quarter mentioning the trend, with a closing summary of direction.

21. Research question: What regulatory or macro event is most often cited
    as a risk factor for [INDUSTRY] in recent public filings? Decision:
    whether to add a new risk to an internal planning document. Sources:
    the filings themselves. Output: the risk factor named, how many filings
    cite it, and one representative quote.

Internal and operational research

22. Research question: What do the top five [TOOL CATEGORY] vendors publish
    about their data retention and deletion policies? Decision: which
    vendor to recommend for a project handling sensitive data. Sources:
    each vendor's own privacy or data-handling documentation. Output: one
    row per vendor: retention period, deletion process, and whether
    exceptions are documented.

23. Research question: What have the last four release notes for [TOOL WE
    USE] changed, and does anything affect how our team currently uses it?
    Decision: whether to update an internal how-to document. Sources: the
    vendor's own changelog or release notes. Output: one line per release,
    flagging anything that looks breaking.

24. Research question: What certifications or audits (for example SOC 2,
    ISO 27001) does [VENDOR] currently hold, and when were they last
    renewed? Decision: whether the vendor clears a security review.
    Sources: the vendor's own trust or security page. Output: certification
    name, issuing body, and most recent renewal or audit date; note
    anything claimed but not evidenced with a document or badge link.

25. Research question: What do published salary or rate benchmarks say
    about [ROLE] in [REGION] as of this year? Decision: setting a budget
    range for a new hire. Sources: named salary-survey publishers, not
    aggregator estimates. Output: the range with its source and survey
    year; note where sources disagree by more than a small margin.

Due diligence on a claim you already hold

26. Research question: Is the statistic "[STATISTIC AS YOU CURRENTLY STATE
    IT]" still accurate, and what is its original source? Decision: whether
    to keep using this number in outward-facing material. Sources: trace
    back to the original study or report, not a page that cites it
    secondhand. Output: original source, publication date, and whether the
    number has been superseded by anything more recent.

27. Research question: Does [VENDOR OR PUBLICATION] still say what we quote
    them as saying about [TOPIC]? Decision: whether a quote in our own
    content needs updating or removing. Sources: the original page, current
    version. Output: the current wording verbatim, and a note if the page
    or the claim no longer exists.

28. Research question: What is the most recent authoritative figure for
    [METRIC], and how does it compare with the number we currently publish?
    Decision: whether to update a stat on our own site. Sources: the metric
    owner's own current publication. Output: current figure, its date, and
    the gap versus our published number.

When Is Deep Research the Wrong Tool?

Four cases where reaching for it costs you time rather than saving it.

Anything time-critical. OpenAI says this plainly: "If time is critical, use search or standard chat for faster responses and save deep research for in-depth analysis." A 20-minute run is the wrong shape for a question you need answered before your next meeting.

Quick lookups or a short back-and-forth. The help center's own framing: "For quick lookups or short conversations, standard chat may be faster." If you expect to refine the question three or four times in a row, a normal chat turn keeps that loop tight; Deep Research's plan-then-run structure adds friction to something that should be fast.

Anything already sitting in documents you hold. If the answer is inside PDFs or notes you already have, and the open web would only add noise, a source-bounded tool built for that job is the better fit: NotebookLM's prompting model works from exactly your own material and nothing else.

Anything needing a guaranteed, narrow source set. Deep Research's site restriction is a real control, but it is a setting inside a general web-research tool, not the tool's whole design. If your job is retrieval-shaped from the start, steering the search first and the generation second, Perplexity's search-shaped prompting is closer to the right mental model, even though the two products document their source controls differently.

Where a Prompt Tool Fits Around Deep Research

We build a prompt generator, not a research or verification service, and this feature is a case where that distinction genuinely matters. A Deep Research prompt is structural: question, decision, source class, recency, exclusions, shape. That structure is identical across a competitive brief, a regulatory question, and a purchase decision, which makes it a natural fit for a saved template with fields you swap rather than rewrite.

What a prompt tool cannot do is the more important half. It cannot make a citation accurate, it cannot reach a page Deep Research's own crawler never saw, and it cannot substitute for opening the source yourself. No tool that writes prompts can. If one habit from this post is worth keeping, it is the one OpenAI's own docs already point toward: read the plan before you approve it, and read the sources-used panel before you trust the report.