Deep Research: the difference between an answer and actual research
A chatbot gives you an answer in two seconds to a question that deserves hours of work. This skill does the opposite: it breaks the question up, searches each part, and shows you where it came from.
New to this topic? Start with Stackable AI Skills Explained: 60+ Modular Capabilities
Why a fast answer is sometimes the problem
Picture a question that genuinely matters. Should we enter this market. What do firms like ours actually charge. Which rules will apply to what we do next year.
Ask that in a normal chat and you get a fluent answer within two seconds. That answer is usually not wrong. It is shallow in a way you cannot see: one round of thinking, no searching, and a tone that sounds equally confident on a guess as on a fact.
For a question with a decision attached, that is the wrong tool. Not because the model falls short, but because you asked a research question of something set to answering mode.
What Deep Research does differently
The skill inverts the order. Instead of answering immediately, your question is taken apart first.
- Splitting. Your question becomes four to eight sub-questions. This is the most important moment in the whole process: a question that is split well is already half answered
- Searching each part. Every sub-question gets its own round of web research, separate from the others
- Synthesising. The results are welded into one report, with numbered sources per claim
That third point is where the difference shows. You do not get text saying "it is safe to assume that", you get text where you can click through from any claim to where it came from. And therefore also see where it did not come from.
When you use it
The honest rule: almost never. And that is the point.
| Situation | What you use |
|---|---|
| "What does this abbreviation mean" | Normal chat, free |
| "What happened in my industry this week" | News and research, one round |
| "Is this article accurate" | Normal chat with web search on |
| "Should we move into this, and what argues against" | Deep Research |
| "What does the market charge for this, with sources" | Deep Research |
Deep Research costs a few thousand credits per report, against 150 for a normal research round. That gap is not arbitrary pricing: multiple search rounds and a considerably heavier model really do go into it. For "just look something up" that is money thrown away, and you get the answer more slowly too.
For a decision with thousands of euros or half a year of work behind it, the same spend is negligible next to what a wrong decision costs.
Before anything starts you are told what it will cost, and with too little balance it does not begin. So a monthly allowance never quietly disappears into a question you asked casually.
The question you ask determines everything
Because the first step is splitting, your phrasing decides how the research runs. A vague question produces vague sub-questions, and after that there is nothing left to save.
What helps: write down what you are going to do with the outcome.
Weaker: "research the market for refrigeration"
Better: "we are considering offering maintenance contracts for refrigeration to hospitality businesses in the south of the Netherlands. What do firms charge, how are those contracts structured, and what are the known problems"
The second version produces sub-questions about pricing, contract structures and risks. The first produces a general market overview you cannot use for anything.
That is the same principle as a good instruction for a scheduled task: the less you leave open, the less gets filled in with assumptions.
The sources are the product
It is tempting to read only the summary. Do not, or at least not only.
The sources sit grouped per sub-question below the report, and that grouping is itself information. If one sub-question has three sources of the same kind, you know that part stands on narrow ground. If another has ten varied sources saying the same thing, you can lean on it harder.
That is a judgement you have to make, and it is deliberately not automated away. A report that asserts its own reliability adds nothing to its reliability.
What you do with the report
Research that stays in your downloads folder produced nothing. Two habits prevent that.
Read the sub-questions first, before the answers. If you do not recognise your own question in them, the research went the wrong way, and starting again with a sharper phrasing is cheaper than reading on.
Then write down in three lines what this means for the decision you had to make. If you cannot, something is missing, and you usually know at that moment exactly which question you should have asked. You ask that as a second round, and it is almost always shorter and sharper than the first.
What it is not good at
- Internal numbers. The skill searches the web. What sits in your own administration is unknown to it
- Very recent news. For "what happened today" one quick round is better and cheaper
- Questions without public sources. A niche nobody publishes about yields a thin report, and no skill can fix that
- Certainty. You get a well-founded overview, not proof
That last one is not hedging. Research reduces uncertainty, it does not remove it, and a tool that promises otherwise is selling you a feeling instead of information.
How this fits the bigger picture
Deep Research is one skill among dozens, and that is exactly the model: you switch on what you need and the rest does not exist for you. How that stacking works is in stackable skills explained.
Where it pairs best is with work you publish afterwards. Research that already sorted out the sub-questions and the sources is a considerably better starting point for an article than an empty document, particularly when you are building authority around a topic.
The skill is available from Growth upwards. Switch it on in the marketplace, then pick it in the chat when you have a question worth it.
