Deep Research is Gemini 3.1 Pro's agentic mode: give it a question and it plans, browses up to hundreds of websites, reasons over what it finds, and returns a multi-page cited report in minutes. The quality of that report is set almost entirely by your prompt. This guide gives you a six-part formula, a weak-versus-strong comparison, and 10 full example prompts you can paste straight in. For a broader library once you know the pattern, keep the best Gemini prompts open in another tab.
What Deep Research does
Deep Research is different from a normal Gemini chat. Instead of answering from memory, it acts as an agent: it drafts a research plan, autonomously visits many sources, and synthesizes them into a structured document with inline citations.
- It plans first. Before browsing, Deep Research shows a research plan you can review and edit — add sub-questions, tighten scope, or name specific sources.
- It browses at scale. A single run can read tens to hundreds of websites, grounded with Google Search, then reason across all of them.
- It cites everything. Reports include sources and citations you can click to verify, plus native visualizations like tables and charts.
- It can use your own data. If you allow it, Deep Research also reasons over your Gmail, Drive, and Chat — otherwise it stays on the public web.
- It scales up. With a 1M-token context, two tiers (Deep Research and Deep Research Max), export to Google Docs, and the option to save a strong prompt as a reusable Gem.
Because it runs autonomously, a vague prompt sends it browsing in the wrong direction for minutes. A precise prompt is what turns all that machinery into a report you can act on.
The six-part prompt formula
Every strong Deep Research prompt covers six things: Objective + Scope + Sources + Comparison criteria + Output format + Recency. That maps cleanly onto Gemini's ICC framework — Intent, Context, Constraints.
| Part | ICC | What it does |
|---|---|---|
| Objective | Intent | The one decision or question the report must answer, stated plainly. |
| Scope | Context | The boundaries — region, industry, time span, what to include and exclude. |
| Sources | Constraints | Which sources to trust: peer-reviewed papers, official filings, reputable outlets; what to avoid. |
| Comparison criteria | Context | The exact dimensions to compare options on (price, accuracy, coverage, risk). |
| Output format | Constraints | Structure you want: an executive summary, comparison tables, a numbered reference list. |
| Recency | Constraints | How current the evidence must be — last 12 months, since 2024, only 2026 data. |
You do not need a heading for each part. Write them as plain sentences and Deep Research will fold them into its plan. The habit that matters is naming all six: skip Sources and it may cite forums; skip Comparison criteria and you get description instead of a decision; skip Recency and it may lean on outdated figures. When your prompt is complete, review the research plan Gemini proposes, edit any weak sub-questions, then approve it.
Weak vs strong example
The same topic, two prompts. The weak one gets a broad, shallow report. The strong one names all six parts and gets a decision-ready one.
Weak:
Research the electric SUV market for me.Strong:
Objective: Help me decide which mid-size electric SUV to buy as a family car in the US. Scope: New models on sale in the US in 2026, priced $40,000-$60,000, with at least 250 miles of EPA range. Sources: Prioritize manufacturer spec pages, EPA and NHTSA data, and established automotive reviewers (Car and Driver, Edmunds, MotorTrend); exclude unverified forum posts. Comparison criteria: EPA range, real-world charging speed, cargo space, safety ratings, warranty, and total 5-year cost of ownership. Output: An executive summary with a clear top pick, a comparison table of all qualifying models across those criteria, and a numbered source list with citations. Recency: Only use data published within the last 12 months.The strong prompt tells Deep Research exactly what to fetch, how to weigh it, and how to present it. That is the whole difference. Use this same shape for the examples below.
Deep Research vs Deep Research Max
Both tiers use the same formula. The choice is depth versus speed.
- Deep Research is the faster tier. Use it for most questions — a market scan, a buying decision, a "catch me up on X" report. It still browses widely and cites sources; it just reasons for less time.
- Deep Research Max is the most comprehensive tier. It browses more sources and reasons longer, which suits high-stakes work: due diligence, a formal literature review, a competitor teardown, or a major purchase. It takes more time in exchange for depth.
A good rule: if a mistake in the report would cost you money or credibility, use Max. Otherwise start with Deep Research and re-run in Max only if the first report feels thin. Either way, review and edit the research plan before it runs, and ask for citations and tables in the prompt so the output is verifiable.
10 example prompts
Each prompt below covers all six parts of the formula. Copy any of them, swap the [bracketed placeholders] for your own details, review the research plan Gemini shows you, and run.
1. Market landscape scan
Objective: Give me a decision-ready overview of the [market/industry, e.g. AI note-taking apps] market so I can decide whether to enter it. Scope: Focus on [region] and the [B2B / B2C] segment; cover market size, growth rate, key players, and pricing models. Sources: Prioritize analyst reports, company filings, and reputable business press; exclude anonymous blog posts. Comparison criteria: Market share, pricing, target customer, and key differentiator per player. Output: An executive summary, a comparison table of the top players, a bulleted list of market trends, and a numbered source list with inline citations. Recency: Only use data from the last 18 months.Best for: founders and product managers sizing up an opportunity.
2. Competitor teardown
Objective: Produce a competitor teardown of [Competitor] so my team can position against them. Scope: Their product, pricing, positioning, recent launches, funding, and public customer sentiment. Sources: Prioritize the company's own site and docs, press releases, funding databases, and reputable review sites (G2, Trustpilot); exclude speculation. Comparison criteria: Compare them to us, [Your company], on features, price, target market, and stated weaknesses. Output: A one-page executive summary, a side-by-side comparison table, a SWOT list, and citations for every claim. Recency: Emphasize the last 12 months; note when a data point is older. Use Deep Research Max for this.Best for: sales and product teams who need a sourced, current view of a rival.
3. Pricing and business-model research
Objective: Recommend a pricing model for my [product type, e.g. SaaS analytics tool] based on how comparable companies price. Scope: Study [5-8] direct and adjacent competitors serving [customer segment]. Sources: Use official pricing pages, published case studies, and reputable industry pricing surveys; exclude outdated cached pages. Comparison criteria: Pricing structure (flat, tiered, usage-based, seat-based), entry price, free tier, and enterprise approach. Output: A comparison table of each competitor's pricing, a summary of common patterns, and 2-3 recommended pricing options for me with pros and cons. Recency: Only current, live pricing verified within the last 3 months.Best for: setting or revising SaaS pricing with evidence.
4. Academic literature review
Objective: Write a literature review on [research question, e.g. the effect of intermittent fasting on insulin sensitivity in adults]. Scope: Focus on human clinical studies and meta-analyses; note sample sizes and study design. Sources: Restrict to peer-reviewed journals and systematic reviews indexed in PubMed or Google Scholar; exclude non-peer-reviewed sources and popular press. Comparison criteria: Compare findings by study design, sample size, effect size, and consistency of results. Output: A structured review with an abstract, a synthesis of findings, a table of key studies (author, year, design, n, main result), noted disagreements, research gaps, and a numbered reference list in APA style. Recency: Prioritize studies from 2020 onward. Use Deep Research Max.Best for: students and researchers who need a cited, structured survey.
5. Technology / method comparison
Objective: Help me choose between [Option A] and [Option B] for [use case, e.g. building a real-time data pipeline]. Scope: Cover both approaches' architecture, performance, cost, and operational complexity at [our scale, e.g. ~1M events/day]. Sources: Prioritize official documentation, engineering blog posts from teams running these in production, and reputable benchmarks; flag vendor-marketing bias. Comparison criteria: Throughput, latency, cost at our scale, learning curve, ecosystem maturity, and failure modes. Output: An executive summary with a recommendation, a comparison table, and a short list of caveats, each with citations. Recency: Only sources from the last 24 months, since these tools change quickly.Best for: engineers making an architecture decision. Pairs well with Gemini prompts for coding.
6. Buying decision (any big purchase)
Objective: Recommend the best [product, e.g. robot vacuum] for my needs and budget. Scope: New models available in [country] under [budget]; my priorities are [priority 1] and [priority 2]. Sources: Prioritize independent review sites, verified buyer reviews, and manufacturer specs; exclude affiliate-driven "best of" listicles with no testing. Comparison criteria: [priority 1], [priority 2], reliability, price, warranty, and running costs. Output: A clear top pick plus one budget and one premium alternative, a comparison table, and a short "who each is for" note, all with citations. Recency: Only models and prices current within the last 6 months.Best for: any considered purchase where reviews are scattered across sites.
7. Local / service provider research
Objective: Shortlist the best [service, e.g. accounting firms for freelancers] for my situation. Scope: Providers serving [city/region] that handle [specific need, e.g. self-employment tax and quarterly filings] for [client type]. Sources: Prioritize provider websites, professional directories, and verified client reviews; exclude paid placements presented as recommendations. Comparison criteria: Services offered, pricing, specialization, availability, and review sentiment. Output: A shortlist of 4-6 providers, a comparison table, and 3-5 questions I should ask each before hiring, with source links. Recency: Only currently operating providers with reviews from the last 12 months.Best for: choosing a professional or service provider with due diligence.
8. Career and salary research
Objective: Help me understand the market for [role, e.g. senior product designer] in [location] so I can plan a job move. Scope: Cover typical salary range, in-demand skills, hiring companies, and career trajectory. Sources: Prioritize reputable salary surveys, job boards, and industry reports; note when data is self-reported. Comparison criteria: Salary by experience level, remote vs on-site pay, and which skills correlate with higher pay. Output: A salary summary with ranges, a table of in-demand skills, a list of active hiring companies, and a short 6-12 month action plan, all cited. Recency: Only data from the last 12 months.Best for: planning a job change with real numbers, not guesses.
9. Personal decision (health / finance, informational)
Objective: Give me a balanced, sourced overview of [decision, e.g. index funds vs high-yield savings for a 2-year goal] so I can make an informed choice. Scope: Explain each option, the trade-offs, and the main risks for someone in [situation]. Sources: Prioritize reputable financial education sites, regulator guidance, and established institutions; exclude promotional content. Comparison criteria: Expected return, risk, liquidity, fees, and tax treatment. Output: A plain-language summary, a comparison table, a "questions to ask a professional" list, and citations. Recency: Use current rates and rules as of 2026. Note clearly that this is general information, not personalized advice.Best for: understanding a personal decision before talking to a professional.
10. Trend and "catch me up" report
Objective: Catch me up on the state of [topic, e.g. AI regulation] so I can brief my team. Scope: What has changed in the last [6 months], who the key players are, and what is likely next. Sources: Prioritize primary documents (official announcements, filings, legislation) and reputable news analysis; exclude opinion pieces presented as fact. Comparison criteria: For each development, note what changed, who is affected, and the level of certainty. Output: A dated timeline of key developments, an executive summary, a "what to watch next" list, and citations for every item. Recency: Focus strictly on the last 6 months.Best for: a fast, sourced briefing on a moving topic. Save this one as a Gem and swap the topic each time. See more finished packs in the best Gemini prompts roundup.
Mistakes to avoid
- Skipping the research plan. Deep Research shows a plan before it browses. Approving it blindly wastes the biggest lever you have — edit weak sub-questions and tighten scope first.
- No source constraints. Without a Sources line, the model may cite forums or SEO listicles. Name the source types you trust and the ones to avoid.
- No comparison criteria. Without explicit criteria, you get description instead of a decision. List the exact dimensions to compare on.
- Forgetting recency. An open-ended prompt can lean on outdated figures. Always give a time window, especially for prices, salaries, and fast-moving tech.
- Not asking for tables and citations. Deep Research can produce native tables and inline citations, but say so in the Output line to get a report you can actually verify.
- Using Max for everything. Max is slower. Reserve it for high-stakes reports; start most questions in standard Deep Research.
- Never exporting or saving. Export strong reports to Google Docs to edit and share, and save repeatable prompts as a Gem so you are not rewriting the formula every time.
Run every prompt through the six-part checklist before you submit, and once a topic recurs, turn it into a Gem. For more targeted packs, see Gemini prompts for data analysis and Gemini prompts for Google Workspace.
Frequently Asked Questions
How long does Gemini Deep Research take?
Most Deep Research runs finish in a few minutes. Deep Research browses tens to hundreds of sources before writing, so a broad or Deep Research Max query can take longer. You can keep working in another tab and Gemini notifies you when the report is ready.
Does Gemini Deep Research cite its sources?
Yes. Every Deep Research report includes inline citations and a list of the sites it consulted, so you can click through and verify each claim. Ask for citations explicitly in your prompt if you want a specific format, such as a numbered reference list.
Can Gemini Deep Research read my Gmail and Drive?
Only if you allow it. By default Deep Research browses the public web. You can connect Gmail, Drive, and Chat so it also reasons over your own documents and messages, but that access is opt-in and controlled in your account settings.
What is the difference between Deep Research and Deep Research Max?
Deep Research is the faster tier and handles most questions well. Deep Research Max is the most comprehensive tier: it browses more sources and reasons longer, which suits high-stakes reports like due diligence, literature reviews, or major purchases. Max takes more time in exchange for depth.
Can I edit the research plan before Gemini runs it?
Yes. Deep Research shows a research plan before it starts browsing. Review it, add or remove sub-questions, tighten the scope, or specify sources, then approve it. Editing the plan is the single fastest way to improve the final report.
Can I export a Deep Research report to Google Docs?
Yes. Every report has an export option that sends the full text, tables, and citations to Google Docs in one click, where you can edit and share it. You can also copy sections directly out of the Gemini response.
Can I save a Deep Research prompt as a Gem?
Yes. If you run the same kind of research repeatedly, save your prompt and instructions as a reusable Gem. The Gem keeps your objective, scope, source rules, and output format, so you only swap in the new topic each time.
How specific should a Deep Research prompt be?
Very specific. State one clear objective, the scope and comparison criteria, which sources to trust, the recency window, and the output format. Vague prompts produce broad, shallow reports; tightly scoped prompts produce focused, decision-ready ones.