TL;DR
Gemini Deep Research runs on Gemini 3.1 Pro (the agent version shipped April 21, 2026) and its real edge is the combination of fresh web data plus one-click export to Google Docs with the citations attached. The single highest-leverage move is editing the research plan before the run starts. Budget about 45 minutes: 5 to write the prompt and edit the plan, 5-10 for the run, the rest to verify citations and rewrite the synthesis in your own voice. You need Google AI Pro ($19.99/mo) or AI Ultra ($99.99/mo) for real briefs; the Free tier caps Deep Research at 5 reports per month and a 32k context window.
What this covers
This is the end-to-end Deep Research workflow as the Gemini app works in June 2026: write the question with a freshness spec, edit the auto-generated research plan, run it, sanity-check the citations against your time window, export to Docs, and spot-check the load-bearing sources.
Key tools and concepts:
- Gemini Deep Research: the multi-step research agent inside the Gemini app, rebuilt on Gemini 3.1 Pro in April 2026. It now generates inline charts and infographics natively and can pull from your Drive and Gmail when you connect Workspace.
- Research plan: before it runs, Gemini shows the angles and sub-questions it intends to chase. You can click Edit plan to rewrite it. This is the spec; the run is the implementation.
- Freshness window: an explicit time bound you put in the prompt (“last 90 days”). Without it, the agent drifts toward older, easier-to-find sources.
- Sources panel: the toggle that decides whether the run reads only Google Search (default), or also your Gmail and Drive.
Who this is for
Workspace-heavy teams (Docs / Sheets / Drive) who need fresh web data baked into research output: product managers tracking competitor moves, market analysts on quarterly cycles, content teams mapping a topic landscape, strategy folks scoping new markets, journalists doing background.
When to reach for it (and when not to)
Reach for it when recency matters (last 30-90 days) and the brief will live in Google Docs and link back to Drive sources. Skip it when you need a deep dive into one specialized topic where the good sources sit behind paywalls — Deep Research’s web focus is a weakness there, and you will see it reach lower in the source-quality stack. For that case, pair Deep Research with a manual literature pass.
Plans, limits, and what you actually get
Deep Research behaves differently per tier. As of June 2026:
| Tier | Price/mo | Context window | Deep Research limit | Model |
|---|---|---|---|---|
| Free | $0 | 32k tokens | ~5 reports / month | Gemini 3.1 Pro (capped runs) |
| Google AI Plus | $7.99 | 128k tokens | Higher than Free | Gemini 3.1 Pro |
| Google AI Pro | $19.99 | 1M tokens (~1,500 pages) | ~20 reports / day | Gemini 3.1 Pro, longer runs |
| Google AI Ultra | $99.99 | 1M tokens | ~200 reports / day | Gemini 3.1 Pro, max compute |
Note: “Gemini Advanced” / “Google One AI Premium” was renamed Google AI Pro in early 2026; it is the same flagship consumer subscription. The free Ultra entry price dropped to $99.99 from its earlier $249.99. Verify current numbers on Google’s subscription page before you quote them, since Google adjusts limits frequently.
There is also a heavier Deep Research Max mode (API and Enterprise, built for asynchronous “run it overnight, brief on my desk by morning” workflows). For interactive in-app work, the standard Deep Research agent is the one you want.
Before you start
- Pick a question with a verifiable answer space. “Top 3 X compared on Y and Z, last 90 days” beats “tell me about X.”
- Decide your freshness window before opening Gemini. Without it the prompt is incomplete.
- If you want it to read Drive or Gmail, connect Workspace first — those sources are off by default in the Sources panel.
- Budget 45 minutes total: 5 min to write the prompt and edit the plan, 5-10 min for the run, the rest to verify citations and rewrite.
Step by step
- Open gemini.google.com, click Tools (or the + / Add files control) in the prompt box, and select Deep Research.
- Open the Sources panel. Google Search is on by default. Turn on Gmail and Drive only if you want internal context in the run; deselect Search if you want a Drive-only brief.
- Type the research question with a freshness spec. For example:
Focus on the last 90 days. Cite sources with publication date. Prefer industry reports and analyst notes over press releases. - Submit. Gemini generates a research plan before doing any work. This is the highest-leverage step: click Edit plan. Remove off-target sub-topics, add the specific angles you care about, name the sources or databases it should prioritize, and tighten vague headings.
- Click Start research. The run typically takes 5-10 minutes (longer on complex multi-angle topics). You can leave the chat; Gemini notifies you when the report is ready. Don’t interrupt unless it stalls past ~25 minutes.
- The report lands in the Canvas panel, often with native charts and infographics inline. Before reading the synthesis, skim the citation list. Any source older than your freshness window is a re-research target — note it, but don’t immediately rerun.
- In Canvas, click Share & export → Export to Docs. The Doc carries the citations and section structure, and is the artifact your team will actually use, comment on, and reuse.
- Manual citation spot-check: open the 3 most load-bearing citations from Drive or the source URL. Confirm each exists and supports the claim Gemini attributed to it. A meaningful share of citations fail this on under-indexed topics; treat those claims as unsupported until you re-verify.
First-run exercise
- Pick a real research need from this week, not a thought experiment.
- Run the full sequence once with your normal phrasing.
- Run it again with the freshness window and source-preference clauses from step 3. Compare citation overlap and date distribution.
- Note which prompt produced the better citation list. That is your template for similar future questions. The intersection of the two citation lists is your high-confidence set.
Quality check
- Did the citation list match your freshness window? If half the sources are older than your bound, the agent drifted — tighten the prompt.
- Did the 3 load-bearing citations pass the spot-check? Below an 80% pass rate, it’s a draft, not a brief.
- Did the synthesis hide disagreements between sources? Watch for “experts agree” — usually a smoothing tell.
- Did you actually export to Docs and use it, or did it sit in the chat? Briefs that live in chat get forgotten by Friday.
How to reuse this workflow
- Save the prompt + freshness window + source-preference combination as a Drive snippet. Reuse it for quarterly updates of the same question.
- For recurring research (weekly competitive scan, monthly market check), keep the prior Doc and diff against the new one — most of the value is in what changed.
- Keep a citation-failure log. Topics where Gemini consistently invents citations tell you the field is paywalled or under-indexed.
- Refresh roughly every two months. The agent’s source bias and plan templates shift as the model rolls.
Common mistakes
- Skipping the research-plan review. The plan is the highest-leverage step; the run is just execution. Default plans are usually too broad.
- Treating “fresh” sources as “good” sources. Recency does not equal quality — press releases and SEO-optimized blogs are fresh too.
- Not exporting to Docs. The chat output gets lost; the Doc gets read and shared.
- Trusting citations without clicking through. Deep Research sometimes produces plausible-looking URLs that don’t support the claim.
- Letting polished prose convince you the analysis is solid. Pretty prose is the most common LLM tell.
Advanced tips
- For technical topics, add
prefer peer-reviewed or official documentation sourcesto the prompt. The model honors source-quality hints meaningfully. - For business topics, add
prefer sources from the last 12 months and prefer named outlets over aggregatorsto avoid stale and laundered content. - Use the Sources panel to do a Drive-first run (internal context) followed by a Search-only run (external context), then synthesize both in Docs.
- Pair with the Gemini Workspace deep workflow: pull internal Drive context first, then Deep Research for external context, then synthesize in Docs.
FAQ
- Free vs Google AI Pro?: Free caps you at about 5 Deep Research reports per month and a 32k context window — fine for quick scans, not real briefs. Google AI Pro ($19.99/mo) gives 1M-token context, roughly 20 reports/day, and longer, higher-quality runs.
- Does it read my Drive and Gmail?: Only if you connect Workspace and turn them on in the Sources panel — both are off by default. When enabled, Drive and Gmail sources appear alongside web sources in the citation list.
- How long does a run take?: Typically 5-10 minutes, longer on complex multi-angle questions. You can leave the chat and get a notification when it’s done.
- Why are some citations from blogs?: When a topic is under-indexed by reputable outlets, Gemini reaches lower in the source-quality stack. Tighten the prompt and name the source types you want.
- What’s Deep Research Max?: A heavier mode (API / Enterprise) using extended test-time compute for exhaustive, asynchronous reports — think nightly due-diligence jobs. The in-app standard agent is the right tool for interactive work.
- Is this better than Perplexity?: Different strengths. Perplexity is faster for ad-hoc lookup; Gemini Deep Research is better for structured briefs that export to Docs with Drive integration.