How to Use AI Deep Research Step by Step: Asking Questions That Get Useful Reports
- Standfirst (first paragraph, distinct from meta): "A good deep research run starts long before you press submit: the clearer and better-scoped your question, the more useful the report that comes back minutes later.
A good deep research run starts long before you press submit: the clearer and better-scoped your question, the more useful the report that comes back minutes later.
Key points
- Open the Deep Research option in ChatGPT, Gemini or Perplexity first, then submit your question — the steps each maker documents are short and similar.
- Write your question like a brief: name the decision, the scope, the constraints and the format you want, instead of asking one broad, blurry question.
- Check the plan and sources where the tool shows them, let the run finish, then verify the report’s key claims before you act on any of them.
Step 0: Decide whether your question is worth a research run
Save rationed research runs for multi-part questions with real consequences.
We have all done it: typed a huge, tangled question into a chatbot and received a huge, tangled answer. Deep research punishes that habit, because a run costs minutes and a slice of a limited allowance — Google limits daily and simultaneous Gemini runs, OpenAI’s published allowances are monthly, and Perplexity uses credits once free usage runs out.
So start with a simple filter, offered here as practical guidance rather than any maker’s rule. If a question can be answered from one page, use ordinary search or chat. If it needs several sources compared or evidence weighed before you spend money or present findings, it is a deep research question. In our view, that distinction separates people who rave about these tools from people who shrug at them.
Step 1: Open Deep Research in your tool of choice
Each maker documents only a few clicks to start, but the button lives in a different place in each app.
In ChatGPT, OpenAI’s instructions are to select “deep research” in the message composer and enter your query. You can attach files or spreadsheets for context. Once a run starts, a sidebar shows the steps and sources used, and you are notified when it is complete.
In Gemini, Google’s help page sets out the fuller sequence. Go to gemini.google.com, click Add Files and then Deep Research, optionally upload files, and optionally click Sources to choose what the research may draw on — Google Search by default, and services such as Gmail or Drive when the Google Workspace app is connected. Enter your question, click Submit, and Gemini creates a research plan. To change it, click Edit plan; then click Start research, and click Open when the report is ready.
In Perplexity, the current help page says to open the plus menu in the search bar, select Deep Research, enter your request and submit it.
Step 2: Write the question like a brief, not a search query
Say what decision the report supports, what is in and out of scope, and how you want the answer organised.
Here is where craft takes over, and everything in this section is our practical advice, not a vendor claim. A weak prompt sounds like a search box: “tell me about home batteries.” A useful one sounds like briefing a researcher: what you are deciding, for whom, where, by when, and what a good answer must contain. Name the constraints that would change the conclusion — budget, country, dates — because an unscoped run will happily research the wrong problem beautifully.
Ask for the working, not just the verdict. A comparison table, a shortlist with reasons, or a section on disagreements between sources makes later checking far easier. The same discipline improves ordinary prompts too; our checklist for writing good ChatGPT prompts is worth five minutes first. And keep the brief focused: two narrow runs beat one sprawling one.
Step 3: Check the plan and the sources before the run settles
Gemini shows a plan you can edit, ChatGPT shows steps and sources as it works, and steering early is cheaper than redoing a report.
Gemini makes this step explicit: a plan appears before the report is generated, and Edit plan exists so you can redirect the research while redirecting is still free. Read that plan like an essay outline from a bright but literal-minded assistant. Has it understood the actual question? Is it about to spend its effort on a side issue you do not care about?
ChatGPT gives you a different window: the sidebar summarises steps and sources as it works, and OpenAI’s February 2026 update says you can track progress in real time, interrupt to refine the run with follow-up prompts or new sources, and restrict searches to trusted sites. Whichever tool you use, the first minute is the cheapest moment to fix a misunderstanding.
Step 4: Let it run, then read the report like an editor
These reports take minutes — 5 to 30 in ChatGPT, usually about 5 to 10 in Gemini — so use the wait, and do not skim the result.
OpenAI says deep research may take 5 to 30 minutes, and you can step away, with a notification when it is done. Google gives Gemini’s usual range as about 5 to 10 minutes, longer for complex reports, and notes you can leave the chat while you wait. Use that wait: decide which two or three claims your decision rests on. Those are the claims you will check first.
When the report arrives, read it like an editor. Start with the conclusion: does it answer the question you set? Then scan for gaps — a missing constraint, an option you expected to see compared. Only then read it properly. A report you merely consume has done half its job; a report you interrogate has done all of it.
Step 5: Run a short verification routine before you act
Open the citations behind the claims that matter, compare one figure against a second source, and treat anything uncited as unproven.
This routine is short enough to do every time, which is why it works. First, open the original source for each load-bearing claim and confirm it says what the report claims. Second, confirm the most important number or date on one independent page. Third, search for the other side: if the report recommends an option, spend two minutes looking for its known problems. Fourth, note anything stated without a citation, and verify it separately or strike it from your decision.
None of this is paranoia: OpenAI warns that deep research can hallucinate and struggle to separate authoritative information from rumours. Our fact-checking checklist for AI answers turns this routine into a habit for any AI output.
What this means
The skill here is not prompting tricks — it is learning to brief, steer and check, the way a good manager handles any researcher.
Most of us were never taught to delegate research. At school you did your own homework; at work you either Googled it yourself or asked a colleague who Googled it for you. So the instinct is to treat this tool like a search box with superpowers and judge it purely on the answer. That undersells it, and in our view it is why so many first attempts disappoint. The report is the end of a process, and you are a participant in that process, not just its audience. It also changes how a disappointing run should feel: often it is not evidence that the tool is weak, but evidence that the brief was vague — and a vague brief is fixable in a way that a weak tool would not be.
Think of the three moments where your input changes everything. The brief decides whether the tool researches your real problem or an easier neighbouring one. The plan check decides whether its effort goes where your decision needs it. The verification decides whether a fluent mistake makes it into your budget or your presentation. None of those moments requires technical skill — just knowing what you need, noticing a misunderstanding, and refusing to pass on an unchecked claim. Anyone who has ever briefed a freelancer already owns these skills.
There is something quietly encouraging in that. Deep research tools reward clarity and scepticism, two qualities that cost nothing and improve with practice. Start with a question that matters to you this week — a purchase, a trip, a decision at work — and run the full routine once. That single, careful run will teach you more than a dozen casual experiments.
Bottom line
Choose the tool, write a brief with scope and constraints, steer the plan early, and never act on a report you have not spot-checked.
The buttons differ slightly between the three tools, but the craft is identical: a focused question, a checked plan, a patient wait and a sceptical read. Master those four habits and deep research becomes the most useful hour-saver in your AI toolkit.
Related Guides
Start with the pillar guide, then build checking and prompting habits.
- AI Deep Research Tools Explained: How ChatGPT, Gemini and Perplexity Research the Web for You
- How to Check AI Answers: A Simple Fact-Checking Checklist
- How to Write Better ChatGPT Prompts: A Practical Checklist
Sources
The official help and announcement pages behind the steps above.
