Skip to content

Prompt Engineering Foundations · Modern assistant features: modes, memory, research and media · lesson 11 of 16 · 12 min

Research with web search, deep research and sources

From "the AI said" to "the source says"

Ordinary chat answers come from what the model learned during training, which can be out of date or simply wrong. When an assistant searches the web, it reads current pages and cites them. When it runs deep research, it plans a series of searches, reads many sources and writes a long, structured report with citations. Assistants such as ChatGPT, Claude, Gemini, Copilot and Perplexity all offer versions of these features, with different names and limits by plan.

Search and citations change your job: instead of trusting the answer, you check the sources behind it.

When to use which

| Situation | Use | |---|---| | A quick current fact (opening hours, a product spec, today's exchange rate) | Web search | | A comparison or overview across many sources (suppliers, tools, regulations) | Deep research | | Explaining a concept that does not change (what is a profit margin?) | Normal chat | | Your own documents | Upload the file instead of searching |

Writing a strong research prompt

A research prompt needs more direction than a chat prompt, because the assistant will make many choices on your behalf.

GOAL: I'm choosing an email marketing tool for a small online shop in
the UK with 3,000 subscribers.

QUESTION: Compare 4 widely used email marketing tools on: monthly price
for 3,000 contacts, ease of use for non-technical users, e-commerce
integrations (Shopify), automation features, and UK/EU data handling.

SOURCES: Prefer the vendors' official pricing and documentation pages
and recent independent reviews. Ignore affiliate listicles.

OUTPUT: A comparison table, then a 150-word recommendation for my
situation, then a list of sources with dates. Flag any price you could
not confirm on an official page.

The parts that matter most: the decision the research supports, the criteria, the type of sources, and what to do with uncertainty.

Checking a research report

Deep research reports look authoritative. They are long, well formatted and full of links. That makes them easy to over-trust. Use this routine:

  1. Spot-check the key claims. Open the sources behind the three claims your decision depends on most.
  2. Check dates. Prices, rules and product features change. Is each source recent enough?
  3. Look for missing viewpoints. Did it only read vendor pages? Did it skip an important option?
  4. Watch for citation mismatch. Sometimes a link is real but does not support the sentence it is attached to.
  5. Ask for the gaps: "What could not be confirmed? Where did sources disagree?"

Worked example: a founder in Dubai

Sara is setting up a small online skincare business in Dubai and asks a deep research tool about labelling and registration requirements for cosmetics. The report is clear and cites several pages. She opens them and finds that two citations are to a general blog, while the official municipality and federal authority pages give more specific requirements. She re-runs the research asking it to "use only official UAE government and regulator sources, and list each requirement with the authority responsible". The second report is shorter but far more useful. She still confirms the final checklist with the relevant authority before launching.

Search for local and multilingual questions

If your question is about a specific country or city, say so. Search results differ by location and language. For markets like Pakistan, the UAE and Saudi Arabia, ask the assistant to also search in Urdu or Arabic where relevant, and to prefer official government sources.

Hands-on: a mini research sprint

1. Choose a real decision you face (a tool, a supplier, a rule you must
   follow).
2. Write a research prompt with GOAL, QUESTION, SOURCES and OUTPUT.
3. Run it with web search or deep research.
4. Open the sources for the three most important claims. Mark each as
   Confirmed, Partly confirmed or Not supported.
5. Send one follow-up asking the assistant to fix anything you could
   not confirm.

Pitfalls

  • Treating the number of citations as a sign of quality.
  • Asking open-ended questions ("tell me about marketing") that produce long, generic reports.
  • Using research for private or internal questions that the web cannot answer.
  • Forgetting that research tools can take several minutes and may count against usage limits.

How to know your research prompt worked

You end with a decision or a short list, every key claim is backed by a source you opened, and the report told you clearly what it could not confirm.

Video lecture: Research with web search, deep research and sources

Research with web search, deep research and sources — From "the AI said" to "the source says" Ordinary chat answers come from what the model learned during training, which can be out of date or simply wrong…

11 chapters · about 7 minutes · captions and full transcript below.

  1. Research with web search and sources
  2. Why sources matter
  3. When to use which
  4. The research prompt
  5. Five-step source check
  6. Worked example: skincare in Dubai
  7. Local and multilingual research
  8. Practice: a research sprint
  9. Example 1: a visa question
  10. Example 2: choosing a payment gateway (illustrative)
  11. Recap

Lecture transcript

Research with web search and sources

The AI said so is not a source. But the AI found these three official pages, and I checked them, is. In this lecture you will learn the difference between web search and deep research, how to write a research prompt that gets useful results, and a five-step routine for checking the sources behind any report. By the end, you will turn research features from a risk into a genuine time-saver.

Why sources matter

Why does this matter? Because decisions increasingly start with an AI answer, and an answer without sources is just a confident opinion. Prices change. Rules change. Products change. A research answer with citations lets you trace every important claim back to where it came from, in minutes. Think of the assistant as a fast research intern: brilliant at gathering and summarising, but you still read the key pages before you sign anything.

When to use which

Ordinary chat answers come from what the model learned during training, which can be out of date or simply wrong. Web search lets the assistant read current pages and cite them. Deep research goes further: it plans a series of searches, reads many sources over several minutes, and writes a long, structured report with citations. Most major assistants offer versions of both, with different names and limits by plan. Use web search for a quick current fact. Use deep research for comparisons and overviews across many sources. Use normal chat for concepts that do not change. And for your own documents, upload them instead of searching.

The research prompt

A research prompt needs more direction than a chat prompt, because the assistant makes many choices for you. Use four parts. Goal: the decision this supports, like choosing an email marketing tool for a UK shop with three thousand subscribers. Question: the criteria, such as monthly price, ease of use, Shopify integration, automation and UK and EU data handling. Sources: prefer vendors' official pricing pages and recent independent reviews, and ignore affiliate listicles. Output: a comparison table, a short recommendation, a list of sources with dates, and a flag on any price that could not be confirmed on an official page.

Five-step source check

Research reports look authoritative. They are long, tidy and full of links, which makes them easy to over-trust. So here is a five-step check. One: open the sources behind the three claims your decision depends on most. Two: check the dates, because prices, rules and features change. Three: look for missing viewpoints; did it only read vendor pages, or skip an important option? Four: watch for citation mismatch, where a link is real but does not support the sentence it is attached to. Five: ask the assistant what it could not confirm and where sources disagreed.

Worked example: skincare in Dubai

Here is a real-world style example. Sara is setting up a small skincare business in Dubai and asks a deep research tool about labelling and registration requirements for cosmetics. The first report cites several pages, but when she opens them, two are general blogs, while the official authority pages give more specific requirements. So she re-runs it: use only official UAE government and regulator sources, and list each requirement with the authority responsible. The second report is shorter and far more useful. She still confirms the final checklist with the authority before launching.

Local and multilingual research

If your question is about a specific country or city, say so, because search results differ by location and language. For markets like Pakistan, the UAE and Saudi Arabia, ask the assistant to also search in Urdu or Arabic where relevant, and to prefer official government sources. And avoid open-ended requests like tell me about marketing; they produce long, generic reports that waste your time and your usage limits.

Practice: a research sprint

Here is a quick practice sprint. Choose a real decision, such as picking a tool, a supplier, or checking a rule you must follow. Write a research prompt with goal, question, sources and output. Run it with web search or deep research. Then open the sources for the three most important claims, and mark each one confirmed, partly confirmed, or not supported. Finally, send one follow-up asking the assistant to fix anything you could not confirm. Fifteen minutes, and you will trust your research far more, for the right reasons.

Example 1: a visa question

A simple example first. You want to know whether you need a visa to visit a friend in another country next month. Weak prompt: do I need a visa? Strong prompt: I hold a Pakistani passport and plan a ten-day tourist visit to the UAE next month. Using official government sources only, tell me whether I need a visa, how to apply, and the processing time. Cite each source and say if anything could not be confirmed. Then open the official page it cites. Visa rules change, and they differ by passport, so the source matters far more than the summary.

Example 2: choosing a payment gateway (illustrative)

Now a business scenario, with illustrative numbers. Omar is choosing a payment gateway for a new online store in Saudi Arabia, expecting around five hundred orders a month. He runs deep research with a clear goal: compare four gateways on setup fees, per-transaction fees, local card support including mada, payout times and integration with his store platform. Prefer official pricing pages. Output a table and flag any fee that could not be confirmed. The report lists fees for all four, but two are marked unconfirmed. He opens the official pages, finds one fee has changed, and corrects it. At five hundred orders a month, even a small difference in per-transaction fees adds up over a year, so those ten minutes of checking are worth it. The numbers are illustrative; the habit is essential.

Recap

To recap. Web search for quick current facts, deep research for comparisons, chat for stable concepts, uploads for your own documents. Write research prompts with goal, question, sources and output. And never treat the number of citations as proof: open the sources behind key claims, check dates, and ask what could not be confirmed. Try this now: run a mini research sprint on a real decision, check three key claims against their sources, and send one follow-up to fix what did not hold up. Next module: your library of more than fifty prompt patterns.

Key takeaways

  • Web search suits quick current facts; deep research suits comparisons and overviews across many sources.
  • Strong research prompts state the goal, criteria, preferred sources and output, including how to flag uncertainty.
  • Citations are not proof: open the sources behind key claims and check their dates.
  • Ask the assistant what it could not confirm, and re-run with better source instructions when needed.

Try it

Run a mini research sprint on a real decision: write a GOAL/QUESTION/SOURCES/OUTPUT prompt, check three key claims against their sources, and send one follow-up to fix what you could not confirm.