GEMINI LABJP
NANOLITE — Nano Banana 2 Lite is here: Google's fastest and most cost-efficient Gemini Image model, made for running lightweight image generation cheaplyOMNIFLASH — Gemini Omni Flash is in public preview, a natively multimodal model that lets enterprises and developers build custom, dynamic video workflowsAGENTS — Managed Agents expand with background: true for async server-side runs and polling, remote MCP server integration, and refreshing credentials across interactionsMEMORY — The Memory Bank IngestEvents API is generally available, decoupling event ingestion from memory generation so you can stream content continuouslyTHROUGHPUT — Provisioned Throughput now lets you submit up to seven pending orders for the same model and regionDEPRECATE — Image generation models shut down on August 17, and the Grok 4.1 family on the Gemini Enterprise Agent Platform on August 20NANOLITE — Nano Banana 2 Lite is here: Google's fastest and most cost-efficient Gemini Image model, made for running lightweight image generation cheaplyOMNIFLASH — Gemini Omni Flash is in public preview, a natively multimodal model that lets enterprises and developers build custom, dynamic video workflowsAGENTS — Managed Agents expand with background: true for async server-side runs and polling, remote MCP server integration, and refreshing credentials across interactionsMEMORY — The Memory Bank IngestEvents API is generally available, decoupling event ingestion from memory generation so you can stream content continuouslyTHROUGHPUT — Provisioned Throughput now lets you submit up to seven pending orders for the same model and regionDEPRECATE — Image generation models shut down on August 17, and the Grok 4.1 family on the Gemini Enterprise Agent Platform on August 20
Articles/Advanced
Advanced/2026-05-22Intermediate

Why Gemini Deep Research Gets Stuck for an Hour — Five Causes and the Workaround for Each

Sometimes Gemini Deep Research sits at 'researching…' for 30 minutes, then an hour, then dies silently. After six months of running market and competitive research across six sites as an indie developer at Dolice, I separated this single symptom into five distinct causes — only one of them actually heals with a retry.

gemini-deep-research2troubleshooting82long-running-taskresearch-workflow

You hand a research question to Gemini Deep Research, the screen says "researching…" and an hour later it's still saying the same thing. I have been running market and competitive research through Deep Research dozens of times a month since the early access window, as an indie developer at Dolice operating six sites — and at one point the "still running overnight, nothing returned" cases started multiplying.

The freeze looks like a generic "the server is slow" problem on the surface. After six months of watching it, I'm fairly sure there are actually five different causes hidden behind that single spinner, and only one of them is the kind that clears with a retry. The other four require you to change how you phrased the request, not how long you wait.

Here are the five causes I now recognize, and the move that gets past each one.

The five patterns behind a stuck Deep Research

In roughly 140 freeze incidents I have catalogued over the past six months:

  1. Too many parallel search steps (about 35% of incidents): when the research plan generates more than 30 parallel search steps, the back half tends to never complete
  2. Repeated access to a rate-limiting domain (about 25%): heavy repeated hits to certain SaaS documentation sites and a few news domains get throttled mid-flight, and the run silently stalls
  3. Research plan is too broad to ever close (about 20%): asks like "comprehensively cover the X industry's market trends and players and revenue models and outlook" collapse during the planning step
  4. Attached file analysis times out (about 12%): large PDFs or spreadsheets break the parse step
  5. Recitation block fires (about 8%): the request implicitly asks for long verbatim quotes, the safety system blocks the output, and the run goes silent

Only the first one ever heals on retry. The other four need a different prompt.

Cutting down parallel search step count

This is the most common failure. Deep Research starts by drafting a plan — "I will run these N searches in this order" — and a broad request can produce 40–60 parallel steps. From my logs, anything above 25 steps starts failing at noticeably higher rates.

The fix is to break the request into multiple smaller research jobs.

  • ❌ Too broad: "Investigate the competitive landscape, revenue models, user trends, and outlook for the Japanese wallpaper app market"
  • ✅ Split: "Identify the top 10 wallpaper apps in Japan by download volume and their operators" + separately "Identify the three dominant revenue models in Japan's wallpaper app market and the ARPU range for each"

I now target 20–25 plan steps per research job. That single change eliminated roughly a third of my freezes.

Avoiding repeated-domain throttling

For research that needs heavy reading from one or two domains — a SaaS reference site, a single API documentation tree — runs stall part way. The fix is two-fold:

  • Phrase the request as a comparison across multiple sources rather than "go deep on this one site"
  • When resubmitting a stalled run, explicitly say "skip the domains used previously and find alternatives"

The second instruction alone took my resubmit success rate above 80% in my own logs.

A checklist for "this plan is too broad"

Cause 3 is easy to catch if you look at the plan Deep Research displays before it starts. I now reject the plan and rephrase if any of these are true:

  • More than 8 plan items
  • The original request contains words like "comprehensive", "everything", or "the full picture"
  • Four or more comparison targets

I run monthly competitive analyses for Lacrima and Mystery, the two blog sites I operate alongside the four AI labs. The early form was "research the revenue, traffic, and content strategy of our top 10 competitor sites" — and it froze nearly every time. The replacement is "compare the content strategy of the three Tier 1 competitors" and a separate "compare traffic scale across these five Tier 2 sites". Stuck runs essentially went to zero.

Pre-processing attached files

If your research includes attached files, the parse step can stall on big ones. From my experience, freezing risk rises when a single file is above 10MB or attachments total above 30MB.

  • Slim PDFs to only the chapters that matter, under 5–10MB
  • Convert spreadsheets to a CSV that contains just the columns you actually want analyzed
  • OCR any image-only PDFs into text-PDF first

This matters more when calling Deep Research through the API. The browser version is more tolerant; the API gives up earlier.

Detecting a Recitation block in disguise

Cause 5 is the hardest to spot. Recitation block is the safety system that triggers when Gemini detects it is about to emit a long verbatim quote — and on Deep Research, it can silently end the run without an error message.

You will know it is recitation when:

  • The run sits in "researching…" for ages, then quietly ends with nothing
  • The same request fails at the same place on two or three retries
  • Your prompt contains "quoting the original announcement" or "citing the paper's text" or similar verbatim-quotation language

Switch verbatim phrasing for summary phrasing: "summarize the announcement in three points", "extract the paper's key claims". On my news-monitoring research, that single substitution erased the stalls.

A pragmatic five-minute recovery routine

When it does freeze, here is the order I now follow:

  1. Close the tab and reopen it — the server-side session is often still alive and the result actually shows up
  2. If still nothing, cancel and resubmit the request split into pieces
  3. If partial results were already shown, screenshot them and request the missing pieces separately
  4. If you suspect cause 5 (recitation), strip verbatim-quote language and resubmit

After enough overnight runs that returned nothing, I gave up on the "fire one wide question before bed" pattern. The new pattern is "fire 3–5 narrow questions in sequence". Completion rate is steadily above 95% and the answers are sharper.

Deep Research is powerful, but it fails silently. I hope this helps anyone else who has spent a morning staring at a still-spinning research tab.

Share

Thank You for Reading

Gemini Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • Copy-paste ready implementation code
  • New advanced guides published daily
  • $5/mo or $10 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

Advanced2026-05-05
Gemini 2.5 Pro in Production: The Pitfalls Nobody Talks About
A practical guide to the production-specific problems with Gemini 2.5 Pro—rate limit architecture, Thinking mode cost control, long-context quality management, and response quality diagnostics—with complete code examples.
Advanced2026-05-05
Gemini 3 Deep Think Not Working as Expected: 5 Common Issues and Fixes
Troubleshoot Gemini 3 Deep Think mode issues systematically. From API misconfiguration and timeouts to unexpected costs and missing thought summaries — 5 real-world problems with working code fixes.
Advanced2026-04-14
When Gemini Gems Ignore Your Instructions or Refuse to Save
Gemini Gems ignoring your custom instructions, failing to save, or resetting mid-conversation? This guide covers all 7 common issues with specific causes and fixes, plus a practical template for building Gems that actually work.
📚RECOMMENDED BOOKS
Build a Large Language Model (From Scratch)
Sebastian Raschka
LLM Dev
Prompt Engineering for LLMs
Berryman & Ziegler
Prompting
AI Engineering
Chip Huyen
AI Eng
* Contains affiliate links
See all →