Get a deep research report you can defend

What you will learn

• When deep research is worth the wait, and when it is not
• How to write a question it can answer, and say which sources count
• How to review the research plan before it runs
• How to check the report before you use any number

The idea

Deep research does not answer at once: it searches, reads many pages and writes a report with citations, often over several minutes. Many assistants have a version of it under different names, for example Gemini, ChatGPT, Claude, Perplexity and Microsoft Copilot. Where we have confirmed it, the tool page says where to find it and on which plans.

Use it for questions that need many sources, such as what studies found on a topic. For a quick fact or a definition, a normal question is faster and easier to check.

The habits from Research with AI and check citations still apply. What changes is the size. A long report with many citations looks thorough, so it is easy to trust. It can still be wrong, or cite pages that do not say what it claims, a hallucination. Treat the result as a draft.

Five habits make a report you can defend:

1. Ask a question that can be answered, with a place, a time period and the decision it supports.
2. Say which sources count. Official pages, published studies and statistics offices usually do. Vendor marketing is a claim, not evidence.
3. Review the plan. Some tools show a research plan or ask questions before they start. Read the plan, edit it if needed, then approve it; this is the cheapest moment to change direction.
4. Ask for a table of claims, each with its source and the exact sentence that supports it.
5. Check before you use any number. Open at least three sources yourself, starting with the ones behind the numbers.

Example

Nadia, an office manager, is asked whether air purifiers would reduce sick days. Her director wants one page. She starts a deep research run:

Question: Is there evidence that air purifiers in offices reduce sick days?
Scope: offices and similar workplaces, studies from the last ten years.
Sources that count: peer-reviewed studies, public health agencies.
Sources that do not count: manufacturer websites and product pages.
Output: a short summary, then a table of claims with source, date and the
exact supporting sentence. Say where evidence is weak.

The tool shows a five-step plan. One step is "compare top-rated models and prices". She deletes it, as she is not choosing a product yet, and adds "find studies that measured absence, not only air quality".

The report lists eleven claims. Nadia opens four sources. Two match. One is a manufacturer's page, used despite her rule, so she drops that claim. The fourth is a real study, but it measured particles in the air, not sick days. Her page says what the evidence shows and what it does not, citing only sources she opened.

Practice now

1. Pick a question that needs several sources, not a quick fact.
2. Write it with a place, a time period and the decision it supports.
3. List the sources that count and the ones that do not.
4. Start a deep research run. If the tool shows a plan or asks questions, fix anything that is off.
5. Ask for a table of claims, each with a source and a quoted sentence.
6. Open at least three sources, including the source of every number you want to use.

At the end you should have a checked table: claims you can defend, and claims you dropped.

Copy this template

Research question: [specific question, with place and time period]
Why I need it: [the decision or document this is for]

Sources that count: [for example official statistics, government pages,
peer-reviewed studies]
Sources that do not count: [for example vendor marketing, product pages,
forums, undated blog posts]

Before you start, show me your research plan and any questions.

In the report:
- a short summary of what the evidence shows and where it is weak
- a table: claim, source title, link, date, and the exact supporting sentence
- your own conclusions, kept separate from what the sources say
- where you could not find a source that counts

Asking AI to flag uncertainty does not make its answer correct. Keep checking the parts that matter.

Check the output

• Does the report answer your question, or a broader one?
• Does every claim in the table have a source, a date and a quoted sentence?
• Did it use sources you said do not count? Drop those claims.
• In the sources you opened, does the page say what the table claims, including any conditions?
• Is every number you plan to use one you found on the source page yourself?

Common mistakes

• Using deep research for a quick fact. It takes minutes and leaves you a long report to check.
• Skipping the plan. A wrong direction wastes the whole run.
• Trusting the report because it is long. Many citations do not make the claims right.
• Treating vendor pages as evidence. A product page is the company's claim. Look for an official or independent source.
• Forgetting which sources are switched on. Some tools can search your email or files as well as the web. Check before you run, and follow your organisation's rules.