Not every question deserves the same research process. A quick fact check, a source-backed recommendation, and a decision that depends on several competing sources should not receive the same level of effort by default.
The simplest rule is this: use web search to answer a focused question; use Deep Search when you need to investigate, compare, and defend a conclusion.
Use web search for a focused current question
Web search is the right starting point when you need one current fact, a direct explanation of a page, or a small number of recent sources. Keep the question narrow and ask for links or citations when they matter.
Examples include:
- Confirming an event date, product detail, or company announcement.
- Checking the current wording of a regulation or public policy page.
- Finding the primary source behind a claim you have already seen.
- Explaining a single unfamiliar term or document.
Give the task a boundary: “Find the official announcement and tell me the publication date,” is clearer than “research this company.” If a fact is time-sensitive—prices, news, schedules, rules, or public figures—ask for current sources regardless of the effort mode you choose.
Use Deep Search for a decision with moving parts
Deep Search is better when the work has several phases: gather sources, compare evidence, identify contradictions, and turn the findings into a recommendation. It is useful for a vendor evaluation, research review, market brief, policy comparison, or any written decision that another person will need to inspect.
Start with a compact brief that names the decision, scope, criteria, audience, and deadline. For example:
Compare three approaches to a customer-support rollout. Assess implementation risk, total cost, staffing implications, and evidence for each recommendation. Prefer primary sources, separate facts from recommendations, and produce a decision matrix for an operations lead.
That tells Deep Search not only what to find, but how the result will be judged.
Ask for a reviewable research structure
For a broader task, ask the model to make its structure visible:
- The questions and boundaries it will investigate.
- The sources that support each important claim.
- Conflicting evidence, missing information, and material uncertainty.
- A clear distinction between sourced facts, interpretation, and recommendation.
- A short list of the checks a human should make before acting.
These requests make the final answer easier to evaluate and make follow-up questions more productive. A useful research result should let you trace a recommendation back to the evidence, not merely sound complete.
Choose effort deliberately
| You need… | Start with… |
|---|---|
| One current fact or page explanation | Web search |
| A quick everyday answer | A lighter chat effort |
| A complex question that needs deeper analysis | Expert |
| A broad, iterative, multi-source investigation | Deep Search |
It is fine to begin with web search and escalate. If the first answer reveals competing sources, unclear definitions, or an important decision, turn that finding into a Deep Search brief. Conversely, do not use a broad research process merely to answer a question that has one authoritative page.
A final quality check
Before you use a research answer, check the dates, open the most important sources, and look for claims that are stronger than the evidence behind them. Treat disclosed limitations as useful information. Research is not weaker because it shows what remains unknown; it is more useful because you know where human judgement is still required.
The aim is not to make every task longer. It is to spend more effort only where it improves the decision.
