Deep Research: How to Run It in ChatGPT, Gemini or Claude

Deep research is a mode in an AI assistant that takes one question, plans the work, reads many sources on its own for several minutes and hands back a long report with citations. ChatGPT, Gemini and Claude each have one, and you probably already pay for at least one of them. The method fits in a line: write the question as a brief, review the plan where the assistant shows one, then open the sources behind every claim you intend to use. That last step is yours. No assistant does it for you.

What deep research does that a normal chat does not

A normal chat answers from one or two quick searches. A deep research run works more like a junior analyst with an afternoon: it breaks the question into parts, searches, reads, decides what to look into next and writes up what it found. OpenAI's own rule of thumb is to use search for quick facts and deep research for depth and thoroughness, because the run takes more time to read and analyze many sources.^[1]

So it fits questions with many moving parts. A market you are about to enter. Five vendors you need to shortlist. What the published studies say about a treatment, a material or a pricing model. It does not fit a single fact, a number you could look up in ten seconds, or anything where the answer sits in one document you already have. For the wider picture of where assistants help in research work, see our guide to AI for research.

Starting a run in ChatGPT, Gemini or Claude

ChatGPT and Gemini follow the same pattern: switch the mode on, describe what you need, approve or adjust the plan, wait. Claude is shorter: switch research on and ask. The buttons differ.

ChatGPT

1. Type /Deepresearch in the prompt, or pick Deep research from the tools menu (+) or the sidebar.
2. Optionally open Sites > Manage sites and list the websites to search. You can restrict the run to those domains or only prioritize them.
3. Describe the outcome. ChatGPT may ask clarifying questions, then shows a proposed research plan you can edit.
4. Let it run. You can follow progress and interrupt to refine the focus.
5. Open the finished report: it has a table of contents, a sources used section and an activity history, and downloads as Markdown, Word or PDF.

Usage varies by plan, and a counter in the product shows the tasks you have left. Availability also depends on your country.

Gemini

1. At gemini.google.com, click Add Files, then Deep Research.
2. Optionally click Sources to add Gmail or Drive, which appear only if Google Workspace is connected. Google Search is included by default.
3. Enter the question and submit. Gemini drafts a research plan.
4. Click Edit plan if it missed something, then Start research.
5. When the report is ready, click Open. Share & export sends it to Google Docs.

You need to be 18 or over and signed in. Gemini limits daily research requests and how many run at once; Google AI Pro and Google AI Ultra users get higher limits.^[2]

Claude

1. Make sure web search is turned on. Research does not work without it.
2. Click the + button at the bottom left of the chat, then Research. A blue indicator appears.
3. Ask the question. Claude searches the web and the tools you have connected, such as Gmail, Google Calendar and Google Docs.

Research is available on paid Claude plans (Pro, Max, Ultra, Team or Enterprise). It counts against the same limits as ordinary conversations, though a session can use them up faster.^[3]

Brief it like an analyst, not like a search box

The report can only be as focused as the request. A deep research AI given six words will pick its own scope, its own time frame and its own idea of a good source. Give it a brief instead. Five lines are enough:

1. The decision. What will you do with the answer?
2. The scope. Country, time frame, company size, budget band in words.
3. The sources you trust. Official pages, filings, peer-reviewed work, or named sites.
4. The shape. A comparison table first, then findings, then open questions.
5. The rule on gaps. Ask it to say plainly when it found nothing instead of filling in.

Then, in ChatGPT and Gemini, read the plan before you approve it. This is the easiest correction you will ever make. A plan that skips your second question, or leans on forums when you asked for official pages, produces a polished report on the wrong thing.

How long does deep research take?

Minutes, not seconds. Google says a Gemini report usually takes about 5-10 minutes because it analyzes many sources, and longer for complex topics; you can leave the chat and get a notification when it is done. OpenAI gives no figure, only that a run takes more time than search. Plan around it. Start the run, do something else, come back.

The run is not where your time goes anyway. Checking is. A report with forty citations takes far longer to verify than it took to generate, and that part does not speed up with a higher plan.

A cited report is not a verified report: keep a source log

Citations tell you where the assistant says a claim came from. They do not tell you the page says that, that the page is current, or that it is a source you would accept. Sometimes the link is fine and the number next to it is not. So before any claim from the report goes into a memo, a pitch or a purchase, log it.

Open a Google Sheet with four columns and add a row for every claim you plan to reuse. Say the question was which scheduling apps suit a ten-chair dental practice:

Three rules keep the log honest. Open the page yourself and find the sentence. Prefer the primary source over whoever summarized it. And treat not opened as not known. If a claim matters and its source is weak, send the assistant back with a narrow follow-up, or go read the original.

Three signs you need a deep research tool, not a chat mode

The built-in mode holds up well for a one-off question answered in a document. It strains in three specific situations, and each one points at a different kind of product.

The same research every week. A competitor watch or a monthly market note means retyping the brief and re-checking from zero each time, because a chat run starts only when you start it. Research set up as an agent on a trigger fits this case.

Gumloop suits the operations or marketing person who wants the weekly pull to land in a spreadsheet without opening anything. Its Company Brain search, on Pro and Enterprise, covers the other half of that job: answers with citations from your own connected sources.

The list is long. Comparing forty suppliers or two hundred companies in one chat run asks a single report to give every item the same attention. Manus AI runs Wide Research across a long list by splitting it up.

That design fits analysts and founders building a table of many similar things. It also has an answer for pages behind a login: Browser Operator reads them in your own Chrome or Edge with your signed-in sessions. If what you need is fresh rows from public pages and no analysis, scraping the website is the simpler route.

The report is not the deliverable. Your boss wants slides. The client wants a spreadsheet. Copying a chat report into either is an hour of formatting and a fresh chance to break a number.

Genspark fits the person whose research always ends as something to present. Whatever the format, the source log still applies: a deck built from unverified claims is just a better-looking mistake. The full comparison of what each product covers is in our guide to choosing an AI research assistant.

One question, one run, ten claims opened

Pick a real question you have this week, one with a decision attached. Write the five-line brief, run it in the assistant you already have, and open the sources behind the ten claims you would actually repeat to someone. How many survive tells you how much to trust the next report. If you find yourself running that same brief again next month, or reformatting the result into slides, that is the moment to look at a product built for it.