Gemini Deep Research is supposed to auto-browse dozens of sources and hand you a deep, cited report in 5-10 minutes — but yours reads like the first three paragraphs of Wikipedia: no data, no counterarguments, half the citations are low-credibility blogs. The mode isn’t weak. Your prompt didn’t give it enough direction, or you’re on the fast “Deep Research” agent when the job needed “Deep Research Max.”
Fastest fix: prompt with directives, not questions — give it explicit scope, 3-5 named dimensions, and a source-type filter — then run a per-dimension follow-up. If you have Google AI Pro or Ultra, pick Deep Research Max for anything that needs real depth.
What changed in 2026 (read this first)
On April 21, 2026, Google replaced the old Deep Research with two agents, both running on Gemini 3.1 Pro (the December 2025 version ran on Gemini 3 Pro):
| Agent | Best for | Trade-off |
|---|---|---|
| Deep Research | Fast, interactive reports; lower latency and cost | Less exhaustive synthesis |
| Deep Research Max | Maximum comprehensiveness; overnight / background jobs | Slower, uses more of your daily quota |
Both add MCP source connectors (point Deep Research at Gmail, Drive, NotebookLM, or custom data streams) and native in-line charts and infographics (rendered as HTML or via Nano Banana). Gemini 3.1 Pro scored 85.9 on the BrowseComp retrieval benchmark, over 25 points above Gemini 3 Pro — so if your last shallow report predates late April 2026, re-running it alone may already help. As of June 2026 the choice of agent and the prompt quality are still what separate a thin report from a deep one.
Which bucket are you in?
| Symptom | Likely cause | Jump to |
|---|---|---|
| 5,000-word survey, nothing deep | Scope too wide | Cause 1 |
| One-page answer, no real research | Scope too narrow | Cause 2 |
| No counterview; reads like advocacy | Prompt contains the conclusion | Cause 3 |
| Citations are blogs / marketing / Reddit | No source-type filter | Cause 4 |
| Vague claims, no numbers | No data requirement | Cause 5 |
| Missing the authoritative paywalled analysis | Some domains aren’t crawlable | Cause 6 |
| Stops at the 80% version | No follow-up drills | Cause 7 |
Common causes
1. Scope too wide
Tell me about AI → the agent can’t pick an angle and returns a broad survey, nothing deep.
Tell me about AI in healthcare in 2026 → reasonable.
Compare Gemini 3.1 Pro vs GPT-5.5 on long-context benchmarks → drillable.
How to judge: would your prompt return 100M results on Google? Too wide.
2. Scope too narrow / no research to do
What's the latest GPT-5.5 context window? (you want one fact) → don’t use Deep Research; a regular Gemini search answers in 30 seconds.
Did Sam Altman tweet on March 5, 2024? (single-fact verification) → same.
Deep Research is for “synthesize multiple sources and form a judgment.”
How to judge: Google finds the answer on one page? You don’t need Deep Research.
3. Prompt contains the conclusion
Why is Gemini better than GPT? → the model writes to confirm the bias; no counterview appears.
Correct: What are the strengths and weaknesses of Gemini 3.1 Pro vs GPT-5.5 across reasoning, coding, and multimodal benchmarks?
How to judge: does your prompt contain “why X is better/worse,” “how X failed,” or similar conclusion-implying phrasing?
4. No source-type specification
Research X → the agent grabs any source (blogs, marketing, Reddit) → low-authority output.
Explicitly: Use academic papers (arXiv, Nature), official documentation, and tier-1 tech publications only.
How to judge: are more than 30% of citations medium.com / personal blogs / marketing pages?
5. No data / no counter-argument requirement
Deep Research doesn’t volunteer them. Add both to the prompt:
Include specific numbers, benchmarks, dates.Include counterarguments and dissenting views.
6. Some domains aren’t crawlable
NYT, WSJ, Bloomberg, paywalled journals — Gemini often can’t reach them, so reports lean on free sources. The MCP/Sources connectors help if the content is in your Drive or Gmail, but the open paywalled web is still a gap.
How to judge: citations are mostly open Web, missing the paywalled authoritative analysis you know exists.
7. No iterative follow-ups
First-pass Deep Research is the 80% version. Without “expand on dimension N,” you’ll stay shallow.
Shortest path to fix
Step 1: Pick the right agent
In the Gemini app (gemini.google.com), click the Add files button (the + in the prompt box), then choose Deep Research from the dropdown. If you’re on Google AI Pro or Ultra and the topic is meaty, choose Deep Research Max instead — it invests more compute per report. Before it runs, Gemini shows a research plan; click Edit plan and tighten any vague step.
Step 2: Use a structured prompt template
Research question: <one-sentence question>
Scope: <time, region, dimensions>
Dimensions: <3-5 concrete dimensions>
Sources: <required source types>
Output format: <structure>
Counterarguments: Include opposing views from at least 3 sources.
Data: Include specific numbers, dates, benchmarks where available.Example (bad to good):
# Bad
Research vertical farming
# Good
Research question: Is vertical farming economically viable at commercial scale in 2026?
Scope: Focus on US + Europe, last 3 years, leafy greens and herbs only.
Dimensions:
1. Unit economics (CAPEX, OPEX, breakeven yield)
2. Energy and water inputs vs traditional ag
3. Recent bankruptcies (Plenty, AeroFarms) - root causes
4. Profitable players and their differentiation
Sources: Academic papers (Nature, Cell), industry reports (Pitchbook, CB Insights), SEC filings.
Output: Executive summary (300 words) + dimension-by-dimension analysis with cited data.
Counterarguments: Include views from ag economists who think it's a fad.
Data: Specific CAPEX per kg of produce, energy intensity in kWh/kg.Step 3: Set your sources before you run
Click the Sources button in the Deep Research interface. Google Search is on by default; add Gmail, Drive, NotebookLM notebooks, or uploaded files if your authoritative material lives there. To restrict the report to your own documents, deselect Google Search. (Gmail/Drive sources only appear if your Google Workspace account is connected to Gemini.)
Step 4: Ask for the source list first, then synthesis
Round 1:
For my research question above, first give me ONLY a list of 30-50 sources you plan to use, organized by:
- Academic papers
- Industry reports
- Mainstream press
- Primary sources (SEC filings, patents)
Do NOT write the report yet. Just the source list.After receiving the list:
- Strip low-quality entries (medium.com / SEO marketing).
- Flag the key paywalled sources so you know what to paste in manually later.
- Round 2: synthesize on the filtered list.
Now write the full report using only sources from this filtered list:
<paste filtered list>Step 5: Per-dimension follow-up drills
After the first report, drill each dimension separately:
On Dimension 2 (energy and water inputs), expand into:
- Per-region kWh/kg numbers for the top 5 vertical farms
- Comparison to traditional greenhouse farming
- How energy cost trends (solar, nuclear) affect 5-year outlook
Output 1500 words with 10+ specific numerical claims.One follow-up per dimension yields 3-5x the depth of a single-pass report.
Step 6: Demand counterarguments + uncertainty
What are the strongest counterarguments to my conclusion?
What data would falsify the conclusion if observed?
What are the sources of uncertainty in your analysis?These three questions shift the report from advocacy to analysis.
Step 7: Verify key citations
A small share of Deep Research citations are stale, mis-numbered, or hallucinated. Verify the load-bearing ones:
- Pick the 5 most critical data citations.
- Open the originals and verify the figure.
- Mark broken / wrong-numbered ones.
- Ask Gemini to rewrite that section:
this citation is broken, replace with another verified source.
Step 8: Paste paywalled excerpts manually
For sources Gemini can’t reach (NYT, Bloomberg, paywalled journals), paste excerpts yourself:
Here are key excerpts from sources you couldn't access:
[paste NYT article key paragraphs]
[paste Bloomberg data table]
Incorporate these into your synthesis.Step 9: Upgrade your plan if you keep hitting limits
Deep Research is available on every tier, but the daily quota and depth scale with the plan. As of June 2026 (note: “Gemini Advanced” / “Google One AI Premium” were renamed in early 2026):
| Plan | Price (USD/mo) | Deep Research runs/day |
|---|---|---|
| Free | $0 | included, lowest limit |
| Google AI Plus | $7.99 | ~3/day |
| Google AI Pro | $19.99 | ~20/day |
| Google AI Ultra (5x) | $99.99 | ~75/day |
| Google AI Ultra (20x) | $199.99 | ~200/day |
Ultra also unlocks richer in-line visuals (charts, diagrams, schematics, interactive simulators). Upgrade at one.google.com/about/google-ai.
How to confirm it’s fixed
A genuinely deep report should pass all four:
- Source count and quality — dozens of sources, most from academic / primary / tier-1 press, not blogs.
- Concrete data — specific numbers, dates, and benchmarks, not “many experts believe.”
- Counterarguments present — at least one section gives the opposing view.
- Citations resolve — your 5 spot-checked citations open and match the claimed figure.
If any fail, run the matching follow-up step above rather than re-running the whole report from scratch.
FAQ
Should I use Deep Research or Deep Research Max? Use plain Deep Research for a fast, interactive answer; use Deep Research Max when you want maximum depth and don’t mind waiting (it’s built for asynchronous, background jobs and burns more of your daily quota). Max runs on the same Gemini 3.1 Pro model but spends more compute per report.
Why are half my citations low-quality blogs?
You didn’t filter sources. Add an explicit source-type line to the prompt (Use academic papers, official docs, and tier-1 publications only) and use the Sources button to scope where it looks. Then ask for the source list first so you can strip junk before synthesis.
Can Deep Research read my Gmail, Drive, or NotebookLM? Yes — via the Sources button (MCP connectors). Google Search is on by default; you can add Gmail, Drive, NotebookLM notebooks, or uploaded files, and deselect Google Search to restrict the report to your own material. Workspace sources only appear if your account is connected to Gemini.
Why does the report ignore the NYT / Bloomberg / paywalled study I know exists? Gemini can’t crawl most paywalled sites, so it falls back to free sources. Paste the key excerpts into the chat yourself and tell it to incorporate them (Step 8).
My report is still shallow after a good prompt — now what? Don’t re-run from scratch. Drill each dimension with a separate follow-up (Step 5) and add the counterargument/uncertainty trio (Step 6). One follow-up per dimension typically multiplies depth 3-5x.
Do I have to pay to use Deep Research? No, it’s on the free tier too, but with the lowest daily limit and shallower output. Google AI Pro ($19.99/mo) raises the limit to about 20 runs/day with deeper reports; Ultra goes higher and adds richer visuals.
Prevention
- Keep a reusable Deep Research prompt template (scope + dimensions + source filter + counterargument requirement).
- Default to Deep Research Max for anything that has to be thorough.
- Set the Sources before you run; ask for the source list first, then synthesis.
- Drill per-dimension follow-ups — don’t expect one shot.
- Verify load-bearing citations; a small share are wrong or stale.
Related
Tags: #Gemini #Debug #Troubleshooting