Research you can trust: Deep Research, citations, verification

Yesterday you picked a daily driver. Today you turn your assistants into a research department — quick sourced answers for small questions, full investigations for big ones — and you learn the 60-second habit that keeps a confident mistake from ever reaching your real life.

1Why a brilliant assistant gets facts wrong

Back on Day 1 we met the Confident Intern: brilliant, tireless, well-read — and sometimes confidently wrong. Today is the day we deal with the "confidently wrong" part properly, because today is the day you start asking AI for facts — facts you might act on, spend money on, or repeat to someone you respect.

Here's the first-principles reason an assistant can flub a fact while sounding completely certain. When you ask a question and the assistant does not search the web, it answers from training memory — everything it absorbed when it was built. That memory is enormous, but it has two built-in weaknesses:

• It's compressed. The assistant didn't store the world's text like a library stores books; it absorbed patterns from it, the way you remember the gist of a novel rather than every sentence. Gist is wonderful for explaining, drafting and brainstorming. It is shaky for exact prices, dates, names and numbers — and the assistant will state the shaky version with the same calm fluency as the solid one. That's the confident mistake we named on Day 1.
• It has a cutoff. Every assistant has a knowledge cutoff — the date its training memory ends, after which an un-searched assistant knows nothing. Ask it about anything that changed recently — a price, a law, a product, an opening time — and an answer from memory can be perfectly phrased and quietly out of date.

Neither weakness is a malfunction. It's what answering-from-memory is. The fix isn't a smarter assistant — it's changing where the answer comes from: have the assistant search the live web first, and then have you check the parts that matter. Those are the two skills of today, and together they're the difference between people AI helps and people AI embarrasses.

2Mode 1 — the quick grounded answer: search first, cite sources

The first research mode is the everyday one. A web search (grounded answer) means the assistant searches the live web before answering, reads what it finds, writes you a short answer — and attaches citations: the clickable references behind each claim. Depending on the assistant, citations show up as small numbered chips after a sentence, as link icons, or as a list of sources at the end. They all mean the same thing: "here's where I got this — go look."

All three of your assistants from Day 3 can do this on their free plans — for Claude, web search is its free research mode. Two things to know about how to read what comes back:

• Citations are for clicking, not decorating. A wall of source chips looks trustworthy, and that look is exactly the trap. A citation only means "I read something there" — it does not guarantee the sentence in front of it faithfully reports what the source says. The only way to know is to open it.
• A claim and its source can drift apart. The most common small failure isn't a fabricated source — it's a real source that says something slightly different: "up to 120 km of range" becomes "120 km of range"; "free for the first year" becomes "free". The direction of the drift is usually toward the cleaner, more confident sentence. Your click catches it in seconds.

One more practical note: if you want a grounded answer, say so. The words "search the web" in your prompt are the reliable trigger. If an answer comes back with no sources at all, that's your signal it came from training memory — tell the assistant to search and cite, and ask again in the same thread (iteration, the Day 1 superpower).

3Mode 2 — Deep Research: a researcher, not a search box

The second mode is the one most beginners have never tried, and it's the closest thing in this course to hiring staff. Deep Research is a different way of working: instead of one search and one answer, the assistant runs a multi-step investigation. It drafts a research plan and shows it to you for approval. (In Gemini you approve a written plan; in ChatGPT the same steering moment looks like a few clarifying questions before it starts — answer them, then let it run.) Then it searches, reads dozens of sources, follows leads, cross-checks — and 5 to 15 minutes later hands you a structured, multi-page report with a citation behind every claim. Minutes, not seconds. A report, not a reply.

Where you can use it free (the workflow is what matters; the exact allowances shift): Gemini gives you a small free allowance of Deep Research runs per month — that's this course's default home for it. ChatGPT's free plan includes a similar small allowance of lightweight runs a month (a lighter version of the same idea). Claude's equivalent (called Research) is paid-only — on the free plan, Claude does web search, not Deep Research. Free allowances change often — check the current limits inside the app, and either way the workflow works the same.

So which mode when? Use the simplest sorting rule there is: minutes-questions vs. money-questions. A minutes-question ("is this museum open on Mondays?", "what does this ingredient do?") deserves a quick grounded answer. A money-question — anything where the answer steers a purchase, a plan, a health choice, or your reputation — deserves Deep Research and your full attention. A small monthly allowance is not a limitation; it's a budget that forces good taste. Spend it on questions that matter.

4The Three-Click Check — the habit that makes both modes safe

Now the verification habit itself. Whatever mode produced the answer, before any AI-sourced fact leaves your screen and enters your real life — an email, a purchase, a conversation, a post — it earns the Three-Click Check. This is the course's canonical verification routine, and it takes about a minute:

Notice what the rule of thumb does: it scales the habit instead of making everything paranoid. Dinner-table trivia? Zero clicks, enjoy. A fact going into an email to your boss? The full three. A claim steering real money, your health, or a legal question? Three clicks, a second opinion, and — for the truly big ones — a human professional. Verification isn't distrust of the tool; it's the same thing you'd do with any talented intern's memo before it goes out under your name. The work stays theirs. The judgment stays yours.