Deep research: multi-step investigations with citations

0:00 Deep research

Deep research can do in fifteen minutes what used to take an analyst two days. It also produces long, confident, citation heavy reports that are very easy to trust too much. In this lecture you will learn what deep research does, how to brief it like a professional, and a six point review checklist that turns its output into something you can defend.

0:27 Why deep research needs discipline

Why does deep research need discipline? Because a long, structured report full of citations feels trustworthy, whether or not every claim holds up. The danger is not that it is useless, it is often excellent, but that its polish discourages checking. Think of it like a very thorough intern's first report. Impressive, mostly right, and still something you review before it goes to the board. The value is created in your brief at the start and your review at the end.

1:02 What deep research does

Deep research is ChatGPT's agentic research mode. You give it a question, and it plans, runs many searches, reads and analyses sources, including files and connected apps if you allow it, then writes a long, structured report with citations. It usually takes several minutes and uses more of your allowance than a normal chat. You can often review its plan before it starts, and restrict it to particular websites or sources.

1:33 When to use it

Use it for broad, multi faceted questions. Market and competitor landscapes. Regulatory overviews across several countries. Evidence reviews. And questions like, what is known about this, and where are the gaps? For a quick fact, normal search is faster. And if you already have the sources, upload them instead.

1:54 The brief

Brief it like a professional researcher. State the purpose and the decision, for example, a strategy memo for the CMO of a modest fashion brand choosing which market to prioritise for a creator programme. List numbered questions. Say which sources you prefer, official regulators, platform policy pages, reputable research firms, and flag vendor marketing as lower reliability, with publication dates. Define the output, an executive summary, one section per market, a comparison table, citations for every claim, and a final section splitting well established findings, uncertain findings, and what could not be found.

2:34 Review the plan first

Before it runs, review the research plan it proposes. Add questions it missed, remove ones that do not matter, and narrow the sources if you need to. Two minutes here saves you reading twenty irrelevant pages later. It is the cheapest quality improvement in the whole workflow.

2:54 The six-point review

Then review like a sceptical editor. Spot check citations, especially numbers and regulatory claims. Check the dates. Check source quality, primary and official versus blogs and vendor content. Ask which statements are the model's own inferences. Look for bias and gaps, such as one region dominating. And recompute any derived numbers, like growth rates or market sizes.

3:19 Statistics are claims

Treat every striking statistic as a claim to verify. Trace it to the original publisher, not a blog quoting a blog. Note who produced it and whether they sell something related. Check the sample and method where you can. Then either attribute it clearly in your deliverable, or leave it out.

3:41 Ethics

A word on ethics. Do not use deep research to compile personal information about private individuals, such as profiling a competitor's employees. It raises real privacy and legal concerns. Keep confidential details out of the brief when they are not needed. And respect copyright, summarise and cite rather than republishing large chunks of sources.

4:04 Simple example

A simple example. Ask deep research to compare three project management tools for a ten person agency, using each vendor's official pricing and feature pages, with the date of each source, in a table. When it returns, open one pricing page and check the price and plan limits in the table. If they match, spot check a second. Five minutes of checking, and you have a comparison you can share with confidence, instead of a vendor's marketing claims repeated back to you.

4:40 Worked example: a Riyadh consultant

A strategy consultant in Riyadh runs deep research on e commerce payment preferences in Saudi Arabia, the UAE and Egypt. In review, she finds one headline statistic came from a payment vendor's marketing report, labels it, and replaces it with an independent figure from a central bank publication. Two findings conflict between sources, so she presents both with dates. Her final memo is shorter than the AI's report, but every claim is defensible. Her review took about forty minutes for a report that would have taken days to compile by hand. She logged every correction, one overstated growth rate, one vendor statistic, one outdated regulation page, and shared the log with her manager alongside the memo. That log became part of the firm's process, so the next consultant knows which kinds of claims need the closest checking.

5:40 Try this now

Try this now. Choose a real question from your work that genuinely needs multiple sources, a market, a regulation, a competitor landscape. Write the brief with purpose, numbered questions, source preferences with dates, and the output shape ending in known, uncertain and not found. If your plan includes deep research, run it, editing the plan before it starts. Then apply the six point review and keep a log of every correction you make. The number of corrections tells you how much to trust future reports on similar topics.

6:18 Common mistakes

Four common mistakes. Using deep research for quick facts, which wastes time and allowance. Forwarding the raw report to a client or manager without review. Trusting statistics that trace back to a vendor selling something related. And treating length as a sign of quality. A two page memo where every claim is verified beats a thirty page report nobody has checked.

6:45 Watch me do it, part 1

Let me run the modest fashion brief. I choose deep research from the plus menu and paste the brief, purpose, the CMO's decision about which market to prioritise, four numbered questions, source preferences with dates, and the output shape ending in known, uncertain and not found. Deep research proposes a plan. It includes a step on general fashion market size, which I do not need, so I remove it. It misses how brands pay creators across borders, so I add a step on payment and tax considerations. Then I start it and get on with other work while it researches.

7:28 Watch me do it, part 2

The report is long and impressive, so I apply the six point review. I spot check citations on the regulatory claims first. Two are fine, one paraphrases a rule more strongly than the regulator's page, so I correct it. A striking platform usage statistic traces back to an agency's marketing report, so I label it vendor sourced and look for an independent figure. I ask which statements are inferences, and it lists three, which I mark as such. Then I open canvas and ask for a one page memo using only the verified findings, with citations. Anything I could not verify simply does not appear.

8:14 Recap and next step

Recap. Use deep research for broad questions, brief it with purpose, questions, sources and output, review its plan, and then apply the six point checklist before anything reaches a client. Your next step: write a brief for a real question in your work, run it if your plan allows, and record every correction you made.