Maya had thirty-one tabs open.
I watched her do it over a video call — she was trying to answer one question for a client memo ("is our market actually growing, or just consolidating?") and every search spawned three more. A report here, a contradictory chart there, a Reddit thread she swore had the real number. Two hours in, she had a browser she was afraid to close and a document that was still blank. "I've read everything," she said. "I just can't hold it."
That's the moment this feature is built for. Not the reading — the holding. JustJot's deep research takes a question, plans the sub-questions a careful person would ask, runs them through web search and source-fetching, and hands you back a multi-section article with every claim linked to where it came from — saved as a normal note you can edit, cut, and publish. The browser-with-thirty-one-tabs problem becomes a draft-with-citations problem, which is a problem you can actually finish.
Here's how to use it so the draft is one you trust.
TL;DR
- Deep research turns a question into a sourced first draft, not a final answer. Treat the output as a starting document with its receipts attached.
- The quality of the report is set by the question. A vague prompt gets a vague plan; a scoped question with a real decision behind it gets a sharp one.
- Read it like an editor, not a reader — every claim is linked, so verify the load-bearing ones in one click instead of trusting the summary.
- The output is a normal note, so the research and the writing live in the same place. Cut what you don't need and the citations travel with what you keep.
- Reach for it when the question is bigger than one search — and reach for [chat with your sources](what-is-ai-chat.md) instead when the answer is already somewhere in your library.
What the feature actually does
Most people picture a smarter search box. It isn't. Search hands you ten links and leaves the holding to you — the exact thing that broke Maya. Deep research does the holding: it decomposes your question, gathers, and synthesizes, then shows its work.
| Stage | What happens | What you get |
|---|---|---|
| Plan | Your question is broken into the sub-questions a careful researcher would ask | A visible outline of what's being investigated |
| Gather | Each sub-question runs through web search and source-fetching | A set of real sources, not a model's memory |
| Synthesize | Findings are written up into sections with claims tied to sources | A multi-section article with inline links |
| Cite | Every load-bearing claim links back; sources collect into a numbered list | A reference list you can audit |
| Save | The whole thing lands as an editable note | A draft you own, in your library |
The thing to internalize: the citations are the point. An answer you can't trace is a rumor with good grammar. An answer where every claim links to its source is something you can stand behind in a client memo — which is the difference between Maya's blank document and a finished one.
Write the question so the plan is good
The report is only ever as good as the plan, and the plan is only ever as good as your question. This is the single highest-leverage minute you'll spend, and it's the one most people skip.
A weak question — "tell me about the coffee market" — produces a weak plan, because the system has to guess what you care about. A strong question carries three things the planner can actually use:
The question checklist - Scope — a boundary. Specialty coffee, not all coffee. North American, not global. The narrower the edge, the sharper the sub-questions. - The decision behind it — what you'll do with the answer. "Should we launch a subscription tier?" pulls different sources than "how big is this market?" - The shape of the answer — a number? a comparison? a timeline? Asking for the shape you need keeps the report from sprawling.
Watch the same curiosity get sharper:
- ❌ "Tell me about the coffee market."
- 🟡 "How big is the specialty coffee market in North America?"
- ✅ "Is the North American specialty-coffee subscription market growing fast enough to justify a new tier in the next 18 months — and what's the strongest evidence against that?"
That last version does Maya's two hours of flailing up front, on purpose. It names the boundary, the decision, and — crucially — asks for the counter-evidence, so the plan goes looking for disconfirmation instead of a comfortable yes.
Read the report like an editor, not a reader
The draft arrives, sectioned and cited, and there's a temptation to exhale and treat it as finished. Don't. The report is a confident-sounding first draft, and confidence is not the same as correctness. Your job now flips from reader to editor: you're not absorbing the text, you're interrogating it.
Because every claim is linked, you can audit by importance instead of reading everything again. Sort the claims by how much weight they carry:
| Claim type | Example | How hard to verify |
|---|---|---|
| Load-bearing | The number your whole decision rests on | Open the source. Always. Check it says what the report says. |
| Supporting | Context that shapes the framing | Spot-check a couple. Confirm the source is credible. |
| Decorative | A nice-to-have aside | Skim. Low stakes if it's slightly off. |
The rule that saves you: click the citation on anything you'd repeat out loud. If you'd put a number in the memo, you open its source before you trust it. This is the same discipline as any AI output — there's a fuller method in [how to verify an AI answer](../ai-literacy/how-to-verify-an-ai-answer.md) — but here it's faster, because the link is right there next to the claim.
Maya's instinct was to read for two hours and write nothing. The editor's move is the reverse: skim the structure in five minutes, then spend your real attention on the three claims the decision actually hinges on.
Turn the report into your own document
Here's the part people miss: the report isn't a destination, it's raw material. It saved as a normal note, which means it lives in the same place as your writing — and the citations stay attached to whatever you keep.
A worked example. Maya's report comes back with six sections. Her client memo needs three. So she:
- Deletes the two sections that don't serve the decision. The market-history section is interesting and irrelevant; it goes.
- Keeps the growth-rate section but cuts it to the one chart that matters — and the source link rides along with the sentence she keeps.
- Writes her own opening — the recommendation, in her voice — above the research, because the report gathered evidence but the judgment is hers to make.
- Pulls the counter-evidence into its own short section. The strongest argument against her recommendation, addressed head-on, is what makes the memo credible.
What used to be thirty-one tabs is now a four-paragraph memo where every number links to its source. The research and the writing never left the same note, so nothing got lost in the handoff between "I read it" and "I wrote it" — the exact gap where Maya's first attempt died.
When not to reach for it
Deep research is for questions that are bigger than a single search and point outward — at the web, at sources you don't have yet. It's overkill, and slower, when the answer is already in things you've saved.
The fork: - The answer is out there and spans many sources → deep research. - The answer is already in your notes → [chat with your sources](what-is-ai-chat.md) instead — it's scoped to what you wrote, so it can't drift into things you never said. - You want a quick fact with low stakes → a plain search is fine; don't bring a research pipeline to a one-line question.
Matching the tool to the question is itself a skill. The [feature decision rule](7-justjot-features-worth-a-second-look.md) covers the wider map of what to reach for when.
Common mistakes
- Trusting the summary because it's well-written. Fluency is not evidence. The citations exist so you don't have to take the prose on faith — use them.
- Asking a vague question and blaming the report. A planless question gets a planless answer. Spend the minute on scope, decision, and shape first.
- Skipping the counter-evidence. If your question only invites a yes, the plan only looks for one. Ask for the strongest case against, explicitly.
- Treating the output as final. It's a first draft with receipts. The judgment, the cuts, and the framing are still yours.
- Verifying everything equally. You don't have time, and you don't need to. Audit the load-bearing claims hard; skim the decorative ones.
Summary — and where to start
Maya's problem was never that she couldn't find information. It was that finding it and holding it were two separate jobs, and the second one kept defeating her. Deep research collapses them: a scoped question becomes a sourced draft in one move, and the draft lives in the same note you'll edit and publish.
Start with one question you've been circling — the kind with thirty-one tabs behind it. Spend a real minute scoping it (boundary, decision, shape, and the case against), run it, then read the result like an editor: skim the structure, click the citations on anything load-bearing, cut to what the decision needs, and write your own opening above it.
The next time you'd open thirty-one tabs, open one note instead. To go deeper on judging any AI-generated draft, read [how to verify an AI answer](../ai-literacy/how-to-verify-an-ai-answer.md); to see where this fits among everything else your library can do, see the [features worth a second look](7-justjot-features-worth-a-second-look.md).