You know the drill: you want to understand something properly — say, which laptop to buy, or whether that new neighbourhood is worth moving to — and an hour later you have forty tabs open, three of them playing videos you never asked for, and you still haven’t made a decision.
Both ChatGPT and Gemini have a button for exactly this situation. It’s usually called Deep Research, and it turns the chatbot from a quick-answer machine into something closer to a research assistant: you describe the question, walk away, and come back to a structured, multi-page report with citations. Most people have never pressed it.
A Normal Answer vs. a Research Report
A normal chatbot answer is written from what the model already knows, possibly with a quick web lookup. A Deep Research run is different: the chatbot plans a research strategy, searches the web on its own, opens and reads pages, follows leads from what it finds, and then writes up a report with footnotes and source links. OpenAI describes it as operating “at the level of a research analyst,” and says a single run can take anywhere from 5 to 30 minutes depending on the complexity of the question.
ChatGPT’s version arrived in February 2025; Google’s Gemini got its own Deep Research mode in late 2024. Both now show you their workings as they go — a sidebar of the steps they’re taking and the sources they’re pulling from — and both keep working in the background if you close the tab. You can also attach your own files, like a spreadsheet or PDF, so the report is grounded in your material as well as the web.
Questions That Are Actually Worth the Wait
Deep Research is overkill for “what’s the capital of Peru?” But it earns its keep on questions that are genuinely multi-part:
- Big purchase comparisons. “Compare these three washing machines on reliability, running costs and warranty terms” — the kind of question where every review site has a different answer.
- Getting up to speed on a topic. A new diagnosis, a pension decision, a home renovation: subjects where you need the landscape, not a single fact.
- Trip or move planning. Neighbourhood comparisons, visa requirements, school catchment areas — information scattered across dozens of pages.
- Work summaries. A competitor overview, a market snapshot, background reading before a meeting.
The common thread is breadth. If the answer requires pulling together information from many places, that’s the job this mode was built for.
Questions That Don’t Need It
Save yourself the wait when the question is simple. A quick fact, a translation, a bit of writing help, an opinion — none of these need a thirty-minute research run. Don’t use it for anything time-critical either: if you need the answer in the next two minutes, the ordinary chat is the right tool. And skip it for personal advice or anything where your own judgement matters more than sources — therapy, relationships, big life decisions.
How to Start One in ChatGPT
In ChatGPT’s message box, look for the Deep research option and select it, then type your question as normal — the more specific, the better. Tell it what to prioritise, what to ignore, and what format you want (a comparison table, a plain-language summary, a report with sections).
ChatGPT will often ask you a few clarifying questions first — answer them honestly, because they genuinely shape the research. Then the work starts. You can close the window and come back later; the research continues in the background. Click the progress indicator if you’re curious what it’s doing — you’ll see its searches and reasoning laid out step by step. When it’s done, you get a long report with citations, and you can ask follow-up questions about it like a normal chat.
How to Start One in Gemini
In Gemini, select Deep Research from the model picker or the Tools menu. One nice touch here: Gemini shows you its research plan before it starts, and you can edit it or suggest changes — worth doing, since it steers the whole run.
Gemini also lets you choose where it looks. Alongside the web, it can pull from your Google Drive, Gmail and Google Chat as sources, which is genuinely useful for work questions — ask it to combine your project documents with public information into one overview. The finished report can be exported to Google Docs, and you can even turn it into an AI-narrated audio summary. Free users get a limited number of runs; a Google AI Pro subscription raises the allowance.
The One Habit That Matters: Check the Citations
Here’s the catch, and it’s worth taking seriously: a long report with footnotes looks authoritative, which makes its mistakes harder to spot. OpenAI itself says Deep Research can “sometimes hallucinate” facts and can struggle to tell authoritative information apart from rumours. An academic quoted in The Guardian warned that people may be tempted to adopt these reports verbatim — and that properly checking one can take hours of work.
So treat every deep research report as a first draft, not a verdict. Click through to the key citations and confirm the important claims on the original pages. If a section has no citations, treat it as the chatbot’s synthesis rather than established fact. For low-stakes questions — a holiday plan, a product shortlist — a quick skim is fine. For anything that involves money, health, legal matters or work decisions, verify the parts that matter before acting.
Used that way, it’s one of the most useful buttons in modern AI. The research still takes time — but it’s the computer’s time now, not yours.
Looking to get more out of the AI tools you already use? You’ll find more practical, plain-English guides in our AI Tools section.
