Framing a deep research brief so the report is actually worth something

The deep research modes in the big AI tools all work the same way. You ask a big question, the tool disappears for anywhere between a few minutes and three quarters of an hour, breaks your question down into sub-questions, searches for each one, reads entire pages and finally puts a structured report with footnotes in front of you. Often several thousand words.

What strikes me is that almost every guide on this argues about which tool is best. That is the boring question. The interesting one is why the same mode gives you a report you can use straight away one time, and five thousand words of mush with nothing usable in it the next. The difference is almost always in the brief, not in the tool.

Here are ten steps for the brief. Checking the finished report is deliberately only a side note here, that is what Fact-checking an AI answer is for. This playbook starts one stage earlier.

Step 1, decide whether you need deep research at all

Deep research is expensive, in time and in quota. Every provider caps the number of runs, and those caps keep changing, which is why there is a counter inside the product itself. Look at it before you burn a run on something a normal question would have answered just as well.

The rule of thumb that works for me: if you could find the answer yourself in two or three searches, use the normal chat with web search. Deep research pays off when your question is made up of several sub-questions and you do not even know exactly which ones those are. Market overview, vendor comparison, getting into an unfamiliar field. Not "what does tool X cost".

Step 2, write the question down as a decision, not as a keyword

The most common mistake is a topic instead of a question. "Project management software" is a topic. The tool then has nothing to steer towards and hands you an encyclopedia.

Write down the decision you actually have to make instead. So not "project management software", but "which project management software fits a team of four that mainly runs client projects with fixed deadlines and has been working with shared spreadsheets so far?" That is longer, and that is exactly why it works. Every extra detail narrows the search space.

Step 3, set the boundaries yourself

A research agent takes whatever it finds, including blog posts that are ten years old and sources from markets that have nothing to do with yours. You have to tell it otherwise.

Four details are worth adding almost every time. The time frame ("only sources from the last eighteen months"), the region ("vendors available in the German-speaking market, billing in euros"), the type of source ("prefer vendor documentation and independent tests, no affiliate comparison sites") and the language. The point about affiliate sites is meant seriously, on software and product questions half your source list will consist of them otherwise.

Step 4, say what shape the result should have

If you say nothing, you get running prose. That is the default form, and for a lot of questions it is the wrong one. A known pattern of these modes is that a question whose answer really is a table gets narrated anyway. Six vendors side by side in prose, unreadable.

So write the shape into the brief. "Result as a comparison table with the columns vendor, price per user, GDPR location, limitations, then at most one page of prose with your recommendation." You can also ask for it afterwards, but in the brief it costs you nothing.

Step 5, check the research plan before it runs off

Some tools show you a plan before they start, meaning the sub-questions they intend to work through, and let you edit it. If you are offered that, use it. Those thirty seconds are the single most effective lever in the whole process.

You can see immediately whether the tool understood your question. If four out of six sub-questions miss your actual problem, you fix that now instead of throwing away a report twenty minutes later. Where there is no plan to look at, you can force one by writing into the brief that the sub-questions should be named first.

Step 6, let it run and do something else

Sounds trivial, it is not. Depending on tool and question these runs take a few minutes to just under an hour. Anyone who sits there waiting cancels, restarts with a slightly different wording, and burns two runs for one result.

Write the brief carefully, start it, and put the window away. If you want to work in parallel, this is exactly the point where you kick off the second run from step 9 as well.

Step 7, read the report backwards

When the report arrives, do not start at the top. Reading five thousand words from the beginning is the surest way to end up impressed and none the wiser.

Read the conclusion first. Then check whether the subheadings really cover your question or whether the tool drifted off along the way. Only then do you read the sections that count for your decision. In most reports that is two out of nine. The rest is completeness, not insight.

Step 8, mark the load-bearing claims

In every report there are a few sentences your decision actually rests on. A price, a legal position, a market size figure. Usually three to five of them.

Mark those. Those are the only ones you really verify, by clicking the source behind them and reading whether it says what the report claims. This matters more than it sounds: the more common failure is not the invented source, it is the real source that gets credited with something it never said. So a link that works is not proof yet. Independent benchmarks now measure exactly this citation faithfulness, and no provider gets a hundred percent.

Step 9, run the same question through a second tool

This is the step with the biggest payoff and the one that is skipped most often. Take the same brief, give it to a second provider, and lay the two reports side by side.

Where both say the same thing, you can be fairly relaxed. Where they contradict each other, you have found the spots you need to check yourself in two minutes. You would never notice those divergences in a single report, because every report sounds self-contained and confident on its own. And that is a second known pattern: on topics that are genuinely contested among experts, these modes average over their sources and present a consensus that does not exist in that form.

Step 10, write the final text yourself

The report is raw material. If it turns into a client email, a proposal, a blog article or a decision paper, you write that yourself, based on the parts you verified.

That is not a question of style, it is a question of liability. The moment you pass the text on, you are the one answering for every number in it, not the tool. And practically speaking, the part of the report you really need is almost always short enough that writing it yourself is faster than rewriting theirs.

What comes next

If you want to understand why these reports sound so assured and can still be wrong, the lesson on Spotting hallucinations in level 1 is the right foundation. And if you want to know which of the big tools brings which research capabilities in the first place, have a look at Tools in ChatGPT, Claude, Gemini in level 2. For checking the result systematically, Fact-checking an AI answer remains the direct follow-on to this playbook.

Sources

The deep research features change quickly, above all the quotas and prices. That is why there are deliberately no concrete numbers on that in the text, only the vendor pages that carry the current state.

• Gemini Deep Research, vendor overview: https://gemini.google/overview/deep-research/
• Claude Research, documentation from Anthropic: https://support.claude.com/en/articles/11088861-using-research-on-claude
• For ChatGPT and Perplexity you will find the current feature set and the limits in the product itself, in the plan section of your account. Both providers block automated retrieval of their announcement pages, which is why there is no direct link here.