Gemini Deep Research is advertised as auto-browsing 50-100 sources and producing a deep report in 10-15 minutes — but your report 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. Wrong scope (too wide / too narrow), no source-type filter, leading question.
To get a genuinely deep report, prompt with directives, not questions: explicit scope, dimensions, and source filters.
Common causes
By shallowness:
1. Scope too wide
Tell me about AI → Deep Research can’t pick an angle and gives a 5K-word survey, nothing deep.
Tell me about AI in healthcare in 2026 → reasonable.
Compare Gemini 2.5 Pro vs GPT-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 password for OpenAI's GPT-5? (you want a fact) → don’t use Deep Research; regular search in 30s.
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? → model writes to confirm the bias; no counterview appears.
Correct: What are the strengths and weaknesses of Gemini 2.5 Pro vs GPT-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 → Deep Research 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: > 30% of citations are medium.com / personal blogs / marketing pages?
5. No data / no counter-argument requirement
Deep Research doesn’t volunteer them:
- “Include specific numbers, benchmarks, dates.”
- “Include counterarguments and dissenting views.”
6. Some domains aren’t crawlable
NYT, WSJ, paywalled academic — Gemini can’t reach them, so reports lean on free sources.
How to judge: citations are mostly open Web, missing paywalled authoritative analysis.
7. No iterative follow-ups
First-pass Deep Research is the 80% version. Without “expand on dimension N”, you’ll stay at shallow.
Shortest path to fix
By depth gained:
Step 1: 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 → 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 2: Ask for 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 (medium.com / SEO marketing)
- Flag the key paywalled sources for Gemini
- Round 2: synthesize on the filtered list
Now write the full report using only sources from this filtered list:
{paste filtered list}Step 3: 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 3-5× the depth of a single-pass report.
Step 4: Demand counterarguments + uncertainty
What are the strongest counterarguments to my conclusion?
What data would falsify the conclusion if observed?
What are sources of uncertainty in your analysis?These three questions shift the report from “advocacy” to “analysis”.
Step 5: Verify key citations
5-10% of Deep Research citations are hallucinations or stale. Verify the load-bearing ones:
- Pick the 5 most critical data citations
- Open the originals — verify
- Mark broken / wrong-numbered ones
- Ask Gemini to rewrite that section: “this citation is broken, replace with another verified source”
Step 6: Use Gemini 2.5 Pro + Advanced
Deep Research on Gemini Advanced (paid) runs on Pro, browsing 50%+ more sources at greater depth than free.
Upgrade: one.google.com/about/ai-premium
Step 7: Paste paywalled excerpts manually
Sources Gemini can’t reach (NYT, Bloomberg, paywalled academic) — 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.Prevention
- Keep a “Deep Research prompt template” (scope + dimensions + source filter + counterargument requirement)
- Always ask for source list first, then synthesis
- Drill per-dimension follow-ups; don’t expect one shot
- Verify load-bearing citations — 5-10% are hallucinations
- Upgrade to Gemini Advanced for deeper crawling and longer context
Related
Tags: #Gemini #Debug #Troubleshooting