ChatGPT Deep Research: Let It Read A Hundred Websites So You Don't Have To

Normal ChatGPT answers in ten seconds from whatever it already knows plus a quick search. Deep Research is different. You ask one big question, it goes away for several minutes, reads dozens of pages, compares them, and comes back with a structured report where every claim has a link next to it. It is the closest thing to having a research assistant that most people will ever get, and it is on every plan now, including Free with a small monthly allowance.

I use it for anything where being wrong costs money or time: a purchase over a few hundred dollars, a trip, a company I am about to deal with, a health question I want to bring to a doctor instead of a search engine. Here's where it lives, how to ask so the report is worth reading, and how to read it without trusting it blindly.

In a normal chat, tap the plus button in the message box and choose Deep research, or just type @Deepresearch at the start of your message. If you are on a paid plan and use Work mode, it lives there too now, and you start it the same way. Then write the request, and this is the whole skill: give it the decision you are making, your situation, what a good answer looks like, and what kind of sources you trust.

I am choosing a robot vacuum for a 900 square foot apartment with one long-haired dog and mostly hard floors, budget up to $500. Compare the top five current models on pet hair pickup, hard floor mopping, noise, app reliability and repair costs. Prioritise reviews from the last 12 months and independent test labs over retailer pages. Give me a comparison table and a single recommendation with the reasons, and tell me the one thing people complain about most for each model.

Swap in your own purchase. The shape is what matters: the decision, the constraints, the format you want back, the sources you trust, and how recent they need to be.

In a normal chat, ChatGPT shows you a proposed research plan before it starts, and you can edit it. In Work mode it asks you a few clarifying questions instead. Either way, do not skip this part. Thirty seconds of "focus on X, ignore Y" here saves you a report about the wrong thing, and once it is running you can still interrupt it to change direction or tell it which sources to use.

Plan a six day trip to Lisbon in late October for two adults who like food markets, walking, and one day trip by train. Budget about 250 euros per day excluding flights. Research the best neighbourhoods to stay in and why, a day by day outline, the best reviewed day trip, and the current rules on transport passes and tipping. Use official tourism and transport sites for prices, say what changes with the season, and end with the three things most first-time visitors get wrong.

While it runs you see its progress live: what it is searching, what it is reading. That is worth watching once, because it shows you how the thing actually works, and it stops feeling like magic and starts feeling like a very fast intern.

The report comes back with citations or source links on every section. In a normal chat you can download it as a Word file, a PDF or Markdown, and in Work it lands as an editable document. Before you act on anything in it, click the links behind the two or three claims that matter most. OpenAI's own line is that it is a first draft, not a final source, and that is exactly right.

This is the run I do before any interview or before handing money to a company I have never dealt with.

I have an interview at [COMPANY] next week for a [ROLE] position. Research what the company does, roughly how big it is and how it is funded, its news from the last six months, the recurring themes in employee reviews on Glassdoor and Indeed, and any public complaints or lawsuits. Separate what is confirmed by the company itself from what is only reported elsewhere. Finish with five smart questions I could ask the interviewer.

For a company you are about to buy from, swap the last part for the refund policy, the Better Business Bureau and Trustpilot patterns, and anything a regulator has said about them.

Three tools, three jobs. A quick fact or a price check is a normal chat with search, ten seconds. A comparison you want written up with sources is Deep Research, several minutes, and it costs one of your monthly runs. Anything that needs a login, a form filled in, or a finished file made is Work, which has its own allowance. Free and Go do not get Deep Research inside Work yet, only in normal chat, which is fine for almost everything on this page.

And the health one, which is the use most people are nervous about. Keep it about the condition and the options, not about a specific person's results, and write it for a conversation with a doctor rather than instead of one.

I have been diagnosed with mild sleep apnea and my doctor mentioned a CPAP machine or a mouthguard as options. Using recent guidance from medical bodies and peer-reviewed sources, summarise the main treatment options for mild obstructive sleep apnea, what the evidence says about each, the typical side effects, and where the evidence is weak or disputed. Then give me eight questions to ask at my next appointment. This is to prepare for a conversation with my doctor, not to replace it.

Claude has the same idea under the plus button as Research, on paid plans only. Gemini's Deep Research is free with lower daily limits. Both work, but the report format in ChatGPT is the one I keep coming back to.

The Honest Part

Deep Research is slow on purpose and it eats a monthly allowance, so it is the wrong tool for anything you could answer in a normal chat, and the counter in the app is the only place OpenAI publishes your exact limit now. The reports look authoritative, which is exactly why you check the links, and it cannot see anything behind a login, so it will not read your bank, your booking, or a members-only site. For that, Work.

Pick the next thing you are about to spend real money on and run Step 1 tonight. It will be the most informed purchase you have made in a while, and it will have cost you four minutes.