Research with citations and deep research

0:00 Research with citations

Research used to mean an afternoon of tabs. Twenty browser tabs, half of them ads, and a nagging doubt about whether any of it was current. AI has changed that. Today, an assistant can search the web in seconds, and deep research modes can read dozens of sources and write you a cited report while you get a coffee. But here's the catch. A confident report with links is not the same as a correct report. In this lecture you'll learn which research mode to use when, how to brief a deep research run so it actually answers your question, and how to get sources you can check. You'll see two examples and watch me brief a real research task.

0:52 Why citations matter

Why do citations matter so much? An assistant's built in knowledge comes from training, which has a cut off date, and it can simply be wrong, even about things it sounds sure of. Rules change. Prices change. Companies launch, merge and close. When facts matter, you need sources: links that show where information came from. Citations do two jobs. They let you check claims quickly, and they let you give credit properly in your own work. Here's the key idea for this lesson. You're not asking the AI to be the source. You're asking it to find the sources, and to show its work.

1:37 Four modes

There are four research modes, and choosing the right one is half the skill. Quick answer, with search off, is fine for explaining a concept or brainstorming. Web search, where the assistant searches right now and answers with links, is for current facts and quick checks. ChatGPT, Claude, Gemini, Copilot and Perplexity all do this. Deep research plans many searches, reads lots of sources and writes a long cited report, which takes a few minutes. ChatGPT's deep research, Claude's Research, Gemini Deep Research, Copilot's Researcher agent and Perplexity's Deep Research are examples. And source grounded tools, like Gemini Notebook, which was called NotebookLM, answer only from files you give them. That's the most reliable option when you already have the right documents.

2:30 The research assistant analogy

Think of deep research like hiring a research assistant for an afternoon. If you say, research the snack market, they'll come back with something, but probably not what you needed. If you say, I need to decide whether to launch in the UAE and Saudi Arabia, focus on data since last year, prefer official trade statistics and retailer announcements, and tell me what you couldn't find, they'll come back with something useful. That's the research brief. Question. Why. Scope, meaning region, time frame and exclusions. Preferred sources. And the output you want. One line that people forget, and it's powerful: tell me what you couldn't find. It forces honesty about gaps instead of filling them with guesses.

3:21 Example 1: the tax threshold

A simple example. Omar runs a small online business in Leeds. His sales are growing and he wants to know whether he needs to register for a particular tax. Instead of asking about tax rules, he asks a specific question: what is the current registration threshold for this tax for a UK online business, and from when does it apply? Use official government sources and give dates. The assistant, with web search on, summarizes the rules and links to the official pages. Omar opens them, confirms the threshold and the date, and notices one detail the summary simplified. He doesn't treat the AI as his accountant. He books a call with his accountant and arrives with three precise questions. Twenty minutes of research instead of an afternoon.

4:16 Example 2: the market-entry scan

Now a realistic business case, with illustrative details. A healthy snacks brand in Karachi is considering exporting to the UAE and Saudi Arabia. The founder, Mariam, runs a deep research report with a full brief: the decision it informs, the two countries, sources published since the start of last year, official trade statistics and major retailer announcements preferred, vendor blogs treated as low confidence. The tool shows a research plan first. She notices it's planning to cover the whole Gulf, so she narrows it to the two countries before it starts. The report comes back with linked findings, a table of sources with dates, and a section saying it couldn't find reliable market size figures for her category. That last section saved her from quoting a made up number in her investor update.

5:14 Watch me do it

Let me brief one. My question: what are the current rules for employers in the UAE offering remote work to staff? Why: our HR lead needs a one page overview before a policy meeting. Scope: UAE federal rules, published or updated in the last two years, excluding free zone specific rules for now. Sources: government and official portals first, reputable law firm summaries second, and forums as low confidence. Output: five bullet summary, findings with a link after every fact, disagreements, what you couldn't find, and a source table. I start the run. It shows a plan that includes free zones anyway, so I tell it to drop that. A few minutes later, I've got the report. Now the measure. I pick five facts at random and open their sources. Four hold up perfectly. One link goes to a page that doesn't mention the claim. So that fact is out until I verify it properly.

6:21 No web access? Bring the sources

What if your tool can't search the web, or your organization has turned search off? You can still research well, and sometimes better. Find the reliable sources yourself: the regulator's page, the company's annual report, the official statistics release. Paste or upload them, and ask the assistant to answer only from these sources and to say which source each point comes from. This is often the most reliable approach for important work, because you've chosen the sources and the AI's job is just to read carefully. It's also how source grounded tools like Gemini Notebook work by design.

7:04 Common mistakes

Common mistakes. Vague questions, which get vague answers. Add purpose, region, time frame and source types. Assuming a link proves the claim. A link only shows where the AI says the information came from. You still need to open it, and that's the whole next lesson. Using AI research about laws, tax, health or finance as final advice. Use it to prepare, then check official sources or ask a professional. Skipping the research plan when a tool shows you one. That's your chance to fix the scope before it spends ten minutes going the wrong way. And putting private information into search queries. Search the topic, not the person or the client.

7:53 Recap and try this now

Let's recap. Choose the right mode: quick answer for concepts, web search for current facts, deep research for multi source reports, and source grounded tools for your own documents. Write a proper brief with the question, the purpose, the scope, the preferred sources and the output, and always ask what it couldn't find. Read the research plan if the tool shows one. Here's your try this now. Pick a real question you need answered this month. Use the research brief from the lesson text in whichever deep research tool you have. Then spot check five facts by opening their sources. That five fact check tells you how much you can trust the rest.