Deep Research Prompt Template That Works in Any Tool (2026)

A good deep research prompt answers seven things before the tool starts: what decision the report feeds, the exact questions, the scope (place, dates, who counts), which sources to prefer or ignore, how strict to be about evidence, what the report should look like, and how long it should be. Skip those and you get a polished but generic overview. Fill them in and the same tool returns something you can act on.

The template below works in ChatGPT, Gemini, Claude, Perplexity and Zemith, because it describes the job and not the tool. Further down is the one step that matters most in any of them: reading the research plan before you press start.

Why does a one-line prompt give a generic report?

A deep research tool turns your prompt into a list of searches, reads a few dozen pages, and writes up what it found. With "research the email marketing software market", it has to guess which market, which buyer, which year and what you will do with the answer, so it picks the safest reading and writes a survey of everything. Every guess it makes for you is a place the report can drift from what you needed.

Vendors say much the same. OpenAI's deep research help article advises describing your question, audience, constraints and the output you want, and notes that in Chat the model drafts a research plan you can review and change (as of October 2026). Google's Gemini help page describes the same step: Gemini builds a plan after you submit, and you can click Edit plan before Start research.

The deep research prompt template (copy and paste)

Replace the brackets. Delete any line that doesn't apply, but think about it first, because each line removes a guess.

DECISION
I'm researching this to decide: [the choice or deliverable this feeds,
e.g. "which of three tools to buy for a 10-person team" or "what to
say in a Thursday client briefing"].

QUESTIONS
Answer these, in this order:
1. [Specific question one]
2. [Specific question two]
3. [Specific question three]
If you find something important I didn't ask about, put it in a short
"Also worth knowing" section at the end.

SCOPE
- Place: [country or region]
- Time: only [last 12 months / since January 2026]; flag older material
- Include: [who or what counts]
- Exclude: [what doesn't]

SOURCES
- Prefer: [official sites, filings, peer-reviewed papers, docs, ...]
- Avoid: [affiliate listicles, vendor blogs for neutral claims, ...]
- Where only a vendor source exists, say so.

EVIDENCE RULES
- Cite a source for every number, price, date and quote.
- Label each key claim: confirmed by two or more independent sources,
  one source only, or inference.
- If sources disagree, show both and say which you trust more and why.
- If you can't find something, say "not found" instead of estimating.

REPORT FORMAT
- Start with a five-line answer to my decision.
- Then one section per question, each ending with a one-line takeaway.
- Put any comparison in a table with the date of each figure.
- End with open questions and what I should verify myself.

LENGTH AND VOICE
About [1,500] words, plain English, no filler, no marketing language.

Three parts of the template earn their place more than the rest:

• Decision. It tells the tool what to leave out. A report for choosing a tool looks very different from one for a client briefing, even on the same topic.
• Evidence rules. The "not found" line and the confidence labels are what stop a confident-sounding report from hiding thin evidence. They also give you a list of claims to check.
• Scope with an exclude line. Naming what doesn't count saves more reading time than anything you add.

What does a filled-in example look like?

Here is the template filled in for a marketer comparing email tools. It's an example of the shape, not a finding about any product.

DECISION
I'm choosing an email platform for a 10-person ecommerce brand with
a list of about 40,000 subscribers. I'll decide this week.

QUESTIONS
1. What does each of Mailchimp, Klaviyo and Brevo cost at 40,000
   subscribers sending 4 campaigns a month?
2. Which of them has built-in ecommerce automations (abandoned cart,
   post-purchase) without a paid add-on?
3. What do recent reviews from stores our size say about
   deliverability and support?

SCOPE
- Place: US pricing in USD
- Time: only pricing pages and reviews from the last 12 months
- Exclude: enterprise plans, agencies reselling the tools

SOURCES
- Prefer: each vendor's own pricing page for prices; review sites
  and Reddit threads for experience
- Avoid: "best email tools" listicles with affiliate links

EVIDENCE RULES
- Cite the exact page for every price and date it was checked.
- Label claims: confirmed by two or more sources / one source /
  inference. Say "not found" rather than estimating.

REPORT FORMAT
- Five-line recommendation first, then one section per question
- One pricing table with the date of each price

LENGTH AND VOICE
About 1,200 words, plain English.

Notice how much of it is dull specifics: a subscriber count, a send volume, a currency. Those are the details a tool would otherwise invent.

What should you check in the research plan before it starts?

Some tools show you their plan first and wait. That pause is the cheapest place to fix a report, so use it. Read the plan with five questions:

1. Does the goal match your decision? If it restates your topic and drops the decision, edit it.
2. Does every question you asked appear? Tools merge or skip questions. Add the missing one.
3. Are the searches aimed at the right sources? Look for official pages, filings or papers where you asked for them, and nothing from the exclude list.
4. Is the date range right? A plan that searches without any date limit on a fast-moving topic will pull in old pages.
5. Does the outline fit how you'll use it? If you need a table, make sure the outline has a place for one.

Fix it in the plan, in plain words, instead of waiting for the report and re-running.

How to run this template in Zemith

Zemith's deep research is built around that review step: it always shows you the plan and waits for you. You need the Plus plan (US$15.99 a month, as of October 2026) or Professional (US$29.99). The Free plan is chat only, with no research tool.

1. Start a New chat and pick a paid model that can use tools. Choose the Claude, GPT or Gemini model you like writing with, since it drafts the final report.
2. Paste your filled-in template and ask for deep research in plain words, for example "Do deep research on this brief" at the top. Pick the Deep Research tool if you'd rather select it than ask.
3. Read the plan card. It opens as "Ready for your review" and lists the research questions, each with the evidence it expects to find. Click Edit plan to change the Research goal, the questions (one per line, up to ten searches), the Report outline and the Source strategy. Then click Start research, or Cancel if the scope is wrong.
4. Close the app if you like. The run keeps going in the background. The default depth reads up to 24 pages and stops within ten minutes, and the report appears in the chat as a cited document.

Your template's SOURCES and SCOPE lines carry through: Zemith's research tool takes preferred, allowed and blocked domains and a freshness window (any, day, week, month or year) as source rules, so "only the last 12 months" and "avoid listicles" are things the plan can act on instead of wishes.

You can also draft the brief in a normal chat first. Describe your decision in two sentences, ask the model to turn it into this template and to ask you any missing questions, then send the finished brief to research. It takes a minute and often catches a missing scope line. The deep research feature page shows the whole flow, and marketers can see how it fits a wider workflow on the AI for marketers page.

How much does a deep research run cost in Zemith?

A run is capped at 100,000 credits, and it can cost less. Plus includes 1.5 million credits a month and Professional includes 3 million, so the cap works out to at most 15 runs a month on Plus or 30 on Professional if you did nothing else. Top-ups are available on paid plans if you run out.

That's one reason to write a tight brief: a focused report costs the same as a vague one, and you'll only want to pay for the focused one.

When is deep research the wrong tool?

Use ordinary chat with web search for a quick fact or a single price check. Deep research is slower and costs more, and it earns that when you need many sources compared and cited. It's also the wrong tool when the answer lives in your own files: put those in a Library project and ask questions of them, because a web research run can't read what isn't on the web.

A few honest limits apply to every tool, including this one:

• Reports can still be wrong. Check the claims labeled "one source only" and any price before you act on it.
• Paywalled and login-only pages are out of reach, so a fact that lives behind one may show up as "not found".
• Vendor pages aren't neutral. The template's "say so when only a vendor source exists" line is there for this.
• Cloud tasks don't run deep research. In Zemith, scheduled and background Cloud tasks are a separate feature, so run research in a normal chat.

If you mostly want fast answers with links rather than a long report, a search-first tool such as Perplexity may suit you better; our comparison of Perplexity Max with Zemith Professional covers where each fits.

The bottom line

Write the decision, the questions, the scope, the source rules, the evidence rules, the format and the length, then read the plan before you start. That template is the whole trick, and it works in any deep research tool. If you want the review step built in, with the plan, the edit and the cited report in one chat, Zemith Plus does it for US$15.99 a month.