Research & Deep Dives

AI with web search: current answers, with receipts

Out of the box, a chatbot answers from what it learned during training — which has a cutoff date and no idea what happened this morning. But most major tools now have a web search mode: ask a question, and the AI searches the internet, reads a handful of pages, and writes you an answer with linked sources — little numbered footnotes or link chips you can click.

Some tools switch to search automatically when your question smells current ("today's dollar rate", "is that visa rule still in force?"); others have a search toggle or button you can press. Either way, the sources are the point. An answer with clickable receipts is a different species from an answer produced purely from memory — you can check it.

One tool category is built entirely around this idea: answer engines like Perplexity, where every response arrives as a short summary sitting on top of a stack of citations. If your main use of AI is "look things up and give me sourced answers", these are worth a look — though the search modes inside ChatGPT, Claude, Gemini, and Copilot have largely caught up.

Deep research: the AI goes away and reads for you

Web search answers a question. Deep research mode — now offered under that name or similar in most major tools — does something bigger: you give it a serious question, and it goes away for many minutes, browsing dozens of pages, comparing them, following leads, and finally returns a structured multi-page report with citations throughout.

It's the difference between asking a friend "which washing machine should I buy?" and asking a diligent assistant to spend the afternoon reading reviews, comparing prices across shops, checking warranty terms, and writing it all up. Good uses from ordinary life:

• Big purchases. "Compare inverter fridges available in Bangladesh under 60,000 taka: reliability, electricity cost, after-sales service."
• Market or business overviews. "What does the home-tutoring market in Dhaka look like — typical rates, how tutors find students, what's changing?"
• Comparing institutions. Universities for your child, banks for a savings scheme, hospitals for a planned operation — anywhere the real answer lives across twenty scattered websites.
• Understanding a health condition. An overview of a diagnosis in the family: what it is, standard treatments, questions worth asking the doctor. A briefing to walk in prepared with — never a replacement for the doctor.

Deep research runs usually cost more of your quota than normal chats (free tiers typically allow a few per month), so spend them on questions that deserve an afternoon of reading — not on things a two-second search would settle.

Writing a good research brief

The report can only be as good as the briefing. A one-line question gets you a generic magazine article; a proper brief gets you a report that reads like it was written for you specifically — because it was. A good brief has three parts:

1. What you want to know — the actual question, with your constraints (budget, location, timeline).
2. What you'll decide with it — this is the magic ingredient. "I'm choosing between X and Y next month" focuses the whole report on your decision instead of general background.
3. What to include and exclude — "include running costs and after-sales service; exclude models not sold in Bangladesh; prefer sources from the last two years."

A fantastic first sweep — not a final authority

Let's be precise about what AI research is and isn't, because both halves matter.

What it genuinely does brilliantly: coverage. In twenty minutes it reads more pages than you'd get through in two evenings, notices options you didn't know existed, and organises a messy topic into a structure you can think with. As a first sweep of any unfamiliar territory, it's one of the best research tools ever put in ordinary hands.

What it isn't: a final authority. It can lean on an outdated page, over-trust a promotional website, miss the one local detail that changes everything (that model has no service centre in your city), or mis-summarise a source — as covered above. And it inherits every caution from the hallucinations topic: fluent, confident, occasionally wrong.

So the working rule is simple: let AI do the sweep; you do the verdict. From any report, identify the two or three sources your decision actually rests on — the official fee page, the warranty terms, the medical association's guideline — and read those yourself. You're not redoing the research; you're auditing its load-bearing walls.

Matching the tool to the question

Two prompts to steal

First, a research brief template. Fill the brackets, then send it to your tool's deep research mode (or a web-search chat for lighter questions):

Research question: [what you want to know, with constraints — budget, location, timeframe]

Why I'm asking: I need to decide [the actual decision, and by when]. Focus the report on what helps that decision.

Please include: [e.g. running costs, reliability, after-sales service, local availability]
Please exclude: [e.g. options above my budget, sources older than 2 years, promotional content]

Structure the report as: a short recommendation up front, then the comparison, then the main uncertainties and what I should verify myself.

Second, a source-check prompt — use it on any AI answer whose sources you're about to trust:

About the answer you just gave me:
1. Which 2–3 of your sources does the conclusion most depend on? List them.
2. For each: how current is it, and is it neutral or promotional?
3. What in your answer is directly stated in the sources, and what is your own inference or interpretation?
4. If one of the key sources turned out to be outdated, which part of your answer would change?

Key takeaways

• Web-search mode gives current answers with clickable sources — and the citation being real doesn't guarantee the summary is faithful. Click through on anything that matters.
• Deep research mode turns a serious question into a many-minute browse and a structured, cited report. Spend it on decisions worth an afternoon of reading.
• A good brief has three parts: what you want to know, what you'll decide with it, what to include and exclude.
• AI research is a fantastic first sweep and a poor final authority. Read the 2–3 sources your decision rests on yourself.
• Plain prompting still earns its keep: "standard objections to X", "steelman both sides", "what am I not asking?"
• Health and legal stakes: AI prepares your questions; professionals give the answers.

Once you've seen the AI browse and report, the next leap will feel natural: watching it build working tools from a plain-language description — and, one step further, act on your behalf. If you want a structured way to practise all of this, the 30-Day Plan schedules it week by week.