ChatGPT Deep Research: How to Use It, With Prompts That Get Better Reports

October 5, 2026

Deep research is the mode in ChatGPT that goes away and does the reading for you. You ask a question, it proposes a plan, then it spends somewhere between five and thirty minutes searching the web and your files before it hands back a long report with citations. Used well, it replaces an afternoon of tabs. Used with a one-line question, it returns a confident summary of whatever was easiest to find.

The difference is almost entirely in what you tell it before it starts. Below: how to start and steer a run, a pattern for the request, five prompts for the jobs people use it for most, and how to check a report before you put your name on it.

How to start a deep research run

There are three ways in, all of which open the same mode:

• Type /Deepresearch at the start of a message.
• Open the + menu in the message box and choose Deep research.
• Pick Deep research in the sidebar.

Every plan has it, in different amounts. OpenAI's pricing page describes it as limited on Free and Go, expanded on Plus and maximum on Pro, and Business, Enterprise and Edu workspaces get it as well, controlled by the admin. OpenAI no longer publishes a number of runs per month. The counter in the product shows how many you have left, and on plans with a monthly allowance it resets 30 days after your first use. Because runs are scarce on the cheaper plans, it pays to get each one right the first time. Prices for each plan are in our ChatGPT review.

What happens during a run

1. It may ask you questions first. ChatGPT checks what you meant before it starts. Answer them properly: this is the cheapest moment to fix the scope.
2. It shows a research plan. You can read it and edit it before the run begins. If the plan covers the wrong ground, change it here, not after half an hour.
3. You can watch and steer. Progress shows live, and you can interrupt to narrow the question or change the sources without losing what it has found.
4. You get a report. It opens full screen with a table of contents, the sources it used and the history of what it did, and you can export it as Markdown, Word or PDF.

By default it reads the public web and anything you upload. Under Sites, you can either restrict it to a list of domains or ask it to prioritise them while still searching the rest of the web. Connected apps such as Google Drive or SharePoint can be sources too, and deep research only reads from them; it never writes.

Deep research, search or a normal chat?

OpenAI's own advice matches this: deep research is built for context and depth, and search is the better tool when recency matters most.

How to write a deep research request

A run follows your request literally and fills every gap with its own judgement. OpenAI's guidance for writing these requests comes down to stating the goal, the scope, the timeframe and the format, and marking anything you haven't decided as open instead of letting the model decide it quietly. In practice that is six lines:

1. The question, in one sentence.
2. What it's for: the decision, the meeting or the reader. «To decide whether to switch providers by March» changes what counts as relevant.
3. Scope and timeframe: the market, the countries, the years. «Since January 2025» keeps out a pile of stale articles.
4. Sources: which kinds to prefer (official documents, filings, studies, the vendors' own pages) and which to treat with care (forums, affiliate reviews).
5. The output: sections, tables, length, and the language.
6. How to handle doubt: ask it to separate evidence from claims and say where sources disagree.

Research question: [one sentence]. I need it to [decide / present / write] [what], for [who reads it].
Scope: [market, countries, years]. Leave out [anything off-limits].
Prefer primary sources: [official sites, filings, studies, vendor documentation]. Treat forums and affiliate reviews as signals, not evidence.
Output: a summary of no more than 150 words first, then [sections], with a comparison table for [what]. About [number] words, in [language].
Separate documented facts from claims and estimates, say where sources disagree, and list open questions at the end.

Five deep research prompts you can paste

Each one follows that pattern for a job people use deep research for again and again. Replace the words in square brackets.

1. Market scan

For a first look at a market before a plan, a pitch or a new product.

Research the market for [product or service] in [country or region] since [year]. I'm deciding whether to [enter it / launch a product / invest time in it].
Cover: who buys and why, the main types of competitors with examples, typical prices, how customers find products, and what has changed in the last 12 months.
Prefer industry reports, official statistics and company filings. Label every estimate as an estimate and name its source.
Output: a 150-word summary, then one section per topic, a table of competitors (name, positioning, price range, source), and five open questions I should answer before deciding.

2. Tool or vendor shortlist

For choosing software, a supplier or a service when several look alike.

Compare [3–6 named tools or vendors] for [your use case] at a team of [size] in [country].
For each: price for our size, the features that matter for [use case], limits that would affect us, data and privacy terms, and what current users complain about.
Use each vendor's own pricing and documentation for facts, and reviews only for complaints. Mark anything you could not confirm on the vendor's site.
Output: a comparison table, then a recommendation with the two reasons that decide it and the main risk.

Before running this one, check our AI tools catalog: if the tools you're comparing are AI products, the scores and prices may already be there.

3. Literature review

For students, analysts and anyone who needs what the research actually says.

Review what published research says about [question] since [year].
Prioritize peer-reviewed studies, systematic reviews and official reports. For each key finding, give the study, its year, its sample or method in one line, and a link.
Show where studies disagree and which side has the stronger evidence. Do not include anything you cannot link to.
Output: a summary of the consensus in 150 words, then findings grouped by theme, then gaps in the research.

4. Competitor deep dive

For understanding one company before a meeting, a pitch or a pricing change.

Research [company] for a [sales pitch / partnership / competitive review]. My company: [one line].
Cover: what they sell and to whom, pricing, recent launches and announcements since [date], how they describe themselves, and what customers praise and criticise.
Use their website, press releases and filings for facts; use reviews and social posts only for sentiment, and say so.
Output: a one-page profile, a timeline of the last 12 months, and three openings for us.

5. Check a claim before you repeat it

For a statistic, a quote or a «study shows» you're about to put in a deck.

Check this claim: "[paste the claim]".
Find its original source, not articles that repeat it. Report: who first published it, when, what the original actually says, and whether later sources confirm, qualify or contradict it.
Give a verdict: accurate, accurate but out of date, misquoted, or unsupported, with the reason.

For shorter jobs that don't need a full run, the Research Briefing Analyst works in a normal chat with search on.

Steering the run

• Answer the clarifying questions in full sentences. «Yes, focus on the UK, small businesses only, ignore enterprise vendors» saves a wasted run.
• Edit the plan if it wanders. Delete steps that answer questions you didn't ask and add the one it missed.
• Restrict sites for anything contested. For health, finance or law, add the official sources under Sites and choose to prioritise them.
• Interrupt early, not late. If the first sources it reads are off-topic, stop and redirect it then. Ten minutes in, the report is already shaped.

How to check a deep research report

OpenAI is direct about the limits: deep research can state wrong facts, can fail to tell rumour from an authoritative source, and often sounds more certain than its evidence. The citations are what make it checkable, so check them.

1. Open the sources behind the three claims you'll rely on. Make sure the page exists and says what the report says it does.
2. Look at what kind of sources they are. A report built on vendor blogs and listicles is a summary of marketing.
3. Watch for numbers with no source. Ask: «List every number in the report with the source it came from.»
4. Check the dates. Deep research cares more about context than recency; for anything that changes monthly, confirm with search.
5. Be wary of a report with no sources at all. Since mid-2026 some Plus users have reported runs that finish with zero searches and zero citations. If yours does, don't use it: run it again, and report it to OpenAI support if it repeats.

Restricting a run to sites you trust is also the main defence against planted information. Press reports in 2026 showed false claims posted on public forums being picked up and repeated by chatbots within a day.

Deep research in other assistants

• Gemini has its own Deep Research with a plan you can edit before it starts, and the Ultra plan adds visual reports. See our Gemini review and the Deep Research Brief prompt written for it.
• Claude calls it Research: on paid plans, switch it on from the + menu with web search enabled, and it can also read connected Gmail, Calendar and Docs. More in our Claude review.
• Perplexity offers Deep Research on its paid plans with live progress as it works. See our Perplexity review.

The request pattern above works in all of them. What changes is the source controls and how long a run takes.

Questions people ask

Is ChatGPT deep research free?

The free plan includes a limited number of runs. OpenAI doesn't publish the exact figure, so check the counter in the product. Plus raises the allowance and Pro raises it again.

How long does a deep research run take?

OpenAI says five to thirty minutes, and you get a notification when the report is ready. Longer questions with many sources take longer, and you can leave the page while it works.

Which model does deep research use?

OpenAI moved deep research to GPT-5.2 in February 2026. Its help page now says only that it runs on the latest model by default, with an option to choose a legacy one, so the exact model may have changed since.

Can deep research read my own files?

Yes. Upload files to the chat, or connect apps such as Google Drive or SharePoint. It reads them alongside the web, or instead of the web if you restrict the sources, and it never changes anything in a connected app.