Writing a DeepSearch Brief Grok Can Actually Act On
"Research our top three competitors' pricing" is the kind of DeepSearch request that produces a report you skim once and never open again. Not because DeepSearch is weak, but because the request left every actual decision to Grok: which competitors, which pricing tiers, how recent, compared against what. DeepSearch will fill in every one of those gaps with a reasonable guess, and reasonable guesses rarely match what you were actually trying to find out. The fix isn't a longer prompt. It's a differently structured one.
Why the same words produce different reports
DeepSearch works by breaking your request into sub-questions, running searches against each one, and synthesizing what it finds into a single write-up. A vague brief gives it vague sub-questions to work from, and the report inherits that vagueness, technically responsive to what you asked, but not useful for what you needed. A specific brief gives DeepSearch a scope, a comparison structure, and a definition of what "done" looks like, and the resulting report reads like something a person who understood the assignment wrote.
The gap between those two outcomes isn't about DeepSearch's capability. It's entirely about how much of the actual decision you handed over versus kept for yourself.
Plan
Sub-questionsYour brief is split into the questions DeepSearch will actually go and answer. Whatever you left unstated becomes a guess baked into this list.
Search
Evidence gatheringIt searches for material on each sub-question. It can only find what the plan asked for, and it treats every source as equally relevant unless you said otherwise.
Write up
SynthesisFindings are assembled into a report. With no requested structure or verdict, it defaults to a balanced, evenly weighted summary.
The important detail is that a vague brief doesn't fail at one stage, it degrades all three, and each stage inherits the previous one's weakness:
| Stage | What a vague brief does to it | What a specific brief does instead |
|---|---|---|
| Plan | Produces broad, generic sub-questions ("what is the company known for?") | Produces sub-questions tied to your decision ("what changed, when, and who is objecting?") |
| Search | Mixes old and new material, and press-release copy with independent reporting | Filters to the window and the source types you named |
| Write up | Balanced summary that avoids a conclusion | Follows your structure and ends with the verdict you asked for |
Bad brief versus good brief: tracking public reaction
The pricing example further down is the classic one. Here is a different kind of research task, where the answer lives partly in what people are saying publicly: the fictional meal-kit company Larkspur Meals is deciding whether to react to a rival's shipping-fee change. Everything below is invented for illustration.
Bad brief: "Find out what people think about FreshCrate's new shipping fees."
That produces something like the following, in the sense of a typical shallow result, not a transcript.
A representative result from the vague brief, illustrated
Every sentence in that summary is defensible and none of it helps Larkspur decide anything. There is no date for the change, no size for the fee, no sense of how many people are reacting, and no way to tell a subscriber from a bystander.
Good brief:
Context: I run marketing at Larkspur Meals, a meal-kit subscription. Our rival FreshCrate changed its shipping fees recently and I need to decide whether we should respond publicly this week. First, establish the facts from FreshCrate's own announcement or help pages: what changed, the old and new fee, and the effective date. Then look at public reaction on X and in news or forum coverage from the two weeks after the announcement. Separate people who say they are current FreshCrate subscribers from general commentary, and tell me roughly whether reaction is a handful of loud posts or a broad pattern. Group the complaints into no more than four themes, with one direct quote per theme and its date. Finish with a short verdict: is there a real opening for us to respond, or is this noise? If the evidence is thin, say so instead of filling the space. Skip listicles about the best meal kits.
”A representative result from the specific brief, illustrated
Four things in the brief did the work. It named the decision, so the plan could aim at "is there an opening" rather than "describe the situation". It separated primary facts from opinion, so the search stage pulled the company's own page first. It asked for a distinction between subscribers and commentators, which stops loud bystanders from counting as evidence. And it permitted the answer "this is thin", which is what keeps the write-up from padding.
The four things a usable brief specifies
Scope: which companies, time period, market, or category, named specifically, not "competitors" or "the market."
Structure: what you want the findings organized by, a table, a ranked list, a specific set of comparison points.
Recency: whether last month's data matters or a year-old report is fine, since DeepSearch will otherwise mix sources of very different ages.
What "done" looks like: a rough sense of length and depth, so it doesn't stop at a surface summary or, just as often, pad a thin finding into three paragraphs.
Worked brief one: competitive pricing for a product manager
Here's the vague version again: "Research our top three competitors' pricing." Now here's a brief that gives DeepSearch something to actually execute against.
I need current pricing research on three specific competitors to our project-management tool: Asana, Monday.com, and ClickUp. For each one, find their current publicly listed pricing tiers, what's included at each tier, and any recent pricing change in the last six months, an increase, a new tier, or a bundling change. Present it as a table: company, tier name, monthly price per seat, and the one or two features that differentiate that tier from the one below it. After the table, add a short paragraph flagging anything that looks like a genuine trend across all three, like a shift toward AI-feature add-ons as a separate paid tier. Prioritize sources from the last three months. Skip general "best project management tools" roundup articles, I want pricing pages and recent pricing-change coverage specifically.
”Notice what this brief does that the one-liner didn't: it names the three companies instead of leaving "top three" to Grok's judgment, it specifies the comparison table's exact columns, it sets a recency window, and it tells DeepSearch which kind of source to deprioritize. Every one of those was a decision the vague version silently pushed onto Grok.
Worked brief two: a finance-side assumption check
DeepSearch is just as useful for a numbers-adjacent research task as it is for competitive intelligence, as long as the brief is equally specific about what would actually change your thinking.
We're modeling a 12% year-over-year growth assumption for a mid-market SaaS renewal rate in our 2027 forecast. Find recent, credible benchmark data on typical SaaS net revenue retention rates for companies in the $10M-$50M ARR range, published within the last year if possible. I want to know whether 12% is conservative, roughly in line, or optimistic compared to what similar companies are actually reporting. Cite the specific source and publication date for each data point, not just a company name, and flag clearly if the benchmark data you find is thin or mostly self-reported by vendors with an incentive to look good, versus independent research.
”This brief does something worth calling out separately: it asks Grok to flag the quality of its own sources, not just report a number. That instruction matters more here than in the pricing example, because a forecast assumption built on a vendor's own marketing claim is a different kind of risk than a pricing table pulled from public pages. Treat any number DeepSearch hands back in a financial context as an input to verify against the original source, not a figure to drop straight into a model.
Tip
Both briefs above ask for a defined output shape, a table or a specific comparison, before asking for analysis. Getting the raw findings organized first makes it much easier to spot what DeepSearch actually found versus where it's stretching a thin result into a confident-sounding paragraph.
When to escalate to a deeper research mode instead
DeepSearch is built for a bounded research task you can describe in a paragraph, the two briefs above are both good fits. Grok has offered a heavier variant, called DeeperSearch when it was introduced, and which research options you see depends on your plan and app version. It exists for the cases where that's not enough: a question that genuinely requires exploring in multiple directions before it's clear what the right sub-questions even are, or a topic where the first round of findings needs to inform a second round of searching before you'd trust the synthesis.
A rough way to tell which one you need: if you can write the brief's structure in advance, the comparison table's columns, the specific entities to research, DeepSearch is the right tool, and a well-structured brief like the ones above will get you most of the value. If you find yourself wanting to say "and then based on what you find, also look into whichever of these seems most significant," that open-ended second layer is what the heavier research options are for, at the cost of a longer run and, depending on your plan, a steeper usage cost.
Common mistake
Reaching for the heaviest research option by default because it sounds more thorough. For a bounded question with a knowable structure, it mostly means waiting longer for a report that a well-written DeepSearch brief would have produced just as well. Save it for the genuinely open-ended cases.
Mistakes that quietly ruin a brief
Naming a topic, not a decision
"Research FreshCrate" gives the plan nothing to aim at. State what you will do differently depending on what the report says.
No source instructions
Without guidance, a company's own marketing and an independent complaint carry equal weight. Say which you want first and which to skip.
Forbidding uncertainty
Demanding a confident answer invites a confident-sounding one built on thin evidence. Ask it to flag weak spots.
Skipping the verification pass
Even a good brief produces a report worth checking. Open the two or three sources your decision leans on.
What to do with the report once it lands
Treat a DeepSearch report the way you'd treat a first draft from a capable junior analyst: read it fully, check the specific claims that matter most to your decision against the actual cited source rather than trusting the summary, and use it as the basis for a follow-up question rather than a final answer. If a source looks thin or the recency window slipped, that's a reasonable thing to send straight back: "the ClickUp pricing you cited looks like it's from an older snapshot, can you check their current pricing page directly and update that row." DeepSearch handles that kind of correction well within the same conversation, since the context of what you already asked for carries forward.