Deep Search is most valuable when the question has moving parts: several sources, conflicting claims, a meaningful decision, or a written outcome that another person needs to review. The quality of that work depends less on a clever one-line prompt than on a brief that makes the decision and the evidence standard explicit.
A research brief is not bureaucratic overhead. It prevents a broad search from becoming a broad answer that sounds useful but cannot be acted on.
Begin with the decision
Start with the choice someone needs to make. “Research AI procurement” is a topic. “Recommend whether the operations team should pilot one of three AI procurement tools this quarter” is a decision.
Name the decision owner and the deadline. This tells the research process how much depth is useful and what kind of result to produce. A board memo, student literature review, and product planning document may cover a similar topic but require very different evidence and tone.
Define the boundaries that matter
Every broad topic needs a few honest limits. State the geography, date range, audience, definitions, and exclusions. If a term has multiple meanings, say which one you mean. If only public information is in scope, say that too.
For example, “compare EU-based software vendors” is still ambiguous. A better instruction says which country group, customer size, contract type, and capabilities matter. Boundaries make the result more useful because they prevent irrelevant facts from taking up the same space as decisive evidence.
Set an evidence standard
Explain what counts as a good source for this task. For a company decision, you may prefer official product pages, pricing pages, contracts, filings, and recent announcements. For academic work, you may prefer primary research, systematic reviews, and the original dataset. For a current event, publication date and event date should both be visible.
Ask Deep Search to distinguish among:
- Directly sourced facts.
- A reasonable inference drawn from sources.
- A recommendation based on the evidence and your stated criteria.
- Missing evidence or a conflict that cannot be resolved from the available material.
This distinction is especially important when a conclusion will be quoted outside the chat. A decision-maker should be able to see which part is evidence and which part is judgement.
Specify the output contract
Describe what a finished result looks like. Do you need an executive summary, a comparison table, a reading list, a chronological timeline, or a set of unanswered questions? State the length, audience, and format. If the answer will be pasted into a briefing, ask for concise headings and source links. If it will be reviewed by analysts, ask for a detailed evidence matrix.
Here is a practical template you can adapt:
Decision: [What should be decided?]
Audience: [Who will use this?]
Scope: [Geography, time range, definitions, exclusions]
Questions: [The 3–5 questions that must be answered]
Criteria: [How options should be evaluated]
Evidence standard: [Preferred source types and freshness]
Output: [Format, length, and required source treatment]
Limitations: [What must be called out rather than guessed]Ask for the research plan to remain visible
For complex work, ask the model to restate the scope in a few lines before it begins, then surface the important trade-offs and gaps in the result. This gives you a chance to correct a misunderstood definition early and makes the final answer easier to audit.
Do not ask for the appearance of certainty. Ask for the opposite: material uncertainty, conflicting evidence, and the exact research question that remains open. Those are often the details that improve a real decision.
Review before you rely on it
Open the sources behind the claims that carry the conclusion. Check dates, definitions, sample sizes, pricing conditions, and any wording that affects legal, financial, or strategic risk. If the brief is being used for a high-impact decision, give it the same human review you would expect from research prepared by a colleague.
A strong Deep Search brief does not make research automatic. It gives the research process a disciplined job: find the relevant evidence, make the reasoning legible, and leave the final judgement to the people accountable for it.
