You wait thirty-five minutes, the spinner finally stops, and the report that comes back reads like a long Wikipedia introduction. Or worse: "Couldn't complete this research" with no further explanation. After running Deep Research for the better part of a year, I have hit both walls many times, and most of them turned out to be avoidable.
Deep Research, when you align with how it actually works, can produce something close to the opening chapter of a master's thesis. But the tool has a few non-obvious behaviors, and missing them costs you about 70% of the depth you could otherwise extract. This article walks through the steps I actually take when a Deep Research run feels too shallow or stalls — in the order I try them.
The official help center only covers fragments of this. The rest is field experience.
Why Reports Feel Shallow — Understanding the Internal Pipeline
Deep Research runs in three stages. Stage one drafts a research plan from your prompt. Stage two crawls and summarizes web sources following that plan. Stage three stitches the findings into a final report.
The crucial point is this: the granularity of stage one sets a hard ceiling on the depth of stage three. If your prompt says "research the latest trends in AI agents," the plan that gets drafted will look like "list of major frameworks / use cases / open challenges" — and no matter how hard stage two works, it cannot go deeper than that scaffolding allows.
In my own observations, prompts under roughly 30 words almost always produce "general overview" reports. Once you get past 100 words of specifics, sections start citing actual product names, version numbers, and benchmark figures. The pattern is consistent enough that I now treat prompt length as a reliable proxy for output depth.
So when Deep Research feels shallow, about 80% of the time the bottleneck is not the model — it is the plan resolution upstream.
Always Edit the Research Plan Before Starting
When you trigger Deep Research, Gemini previews the research plan before crawling. Always edit this plan once before clicking "Start research." This single habit changes the perceived depth of results by roughly 1.5x.
Three checks I run on every plan:
First, I confirm that my actual question is represented as a section in the plan. If the topics I care about are missing, I add them. For "AI agent landscape in 2026," I will manually add lines like "comparison of evaluation benchmarks (HumanEval-style)" or "common failure modes observed in production deployments" rather than letting the plan stay at the surface level.
Second, I explicitly request English-language primary sources. A line like "always reference papers, vendor official blogs, and GitHub README files" measurably shifts the source distribution away from second-hand Japanese summary sites and toward primary documentation.
Third, I specify multiple comparison targets up front. "Research A" produces a meandering report. "Compare A, B, and C across axes X, Y, Z" produces something with structure, tables, and a real conclusion.
Two minutes of plan editing reliably saves you from a 30-minute "redo this completely" later. It is one of the cheapest investments in the entire workflow.
Three Things to Check When Deep Research Stalls
When Deep Research ends with "Couldn't complete this research" or quietly stops mid-run, the cause almost always falls into one of three buckets.
One: blocked source access. If the plan asks Gemini to consult paywalled report sites, specific Twitter threads, or login-gated investor relations pages, those steps can derail the entire run. Removing constraints like "fetch the PDF from this URL" or "summarize this Twitter thread" often unblocks the rerun.
Two: safety filter activation. Medical, legal, and biographical research can trip the safety layer when the prompt asks for definitive judgments. Reframing "give me the medical evidence for X" as "summarize current research and dissenting positions on X" usually goes through cleanly.
Three: long-running job timeout. Deep Research appears to have an internal ceiling around 30–40 minutes. Topics like "evaluate the entire history of the AI industry" never finish. Splitting a topic — say, "the labor-economics shift around the Industrial Revolution" instead of "history of work" — gets you deeper results faster, even though it is two runs instead of one.
When something stalls, my first reaction is to trim the prompt to about 70% of its original scope and loosen the source constraints. That alone resolves roughly seven out of ten failed reruns.
When Sources Skew Toward Japanese Summary Sites
If you are running Deep Research in Japanese (or any non-English language), you may notice the citations cluster around aggregator blogs and news rewrites rather than primary sources. This is the natural intersection of prompt language and Gemini's search preferences.
The fix is direct: append a source-language instruction to the prompt. Something like:
"Prefer English-language primary sources (official blogs, papers, GitHub repositories, IR pages) over Japanese summary sites. Write the final report in Japanese."
In my own measurements, this single instruction shifted the share of English primary sources from about 30% up to 70%. The report itself still arrives in your target language, so reading cost does not go up.
You can amplify the effect by writing numerical priorities into the plan — "cite at least three peer-reviewed papers" or "consult official documentation before secondary commentary." Specific numbers steer the crawler more reliably than abstract preferences.
The structured prompt pattern in our Gemini prompt engineering guide (role, task, constraints, format) translates well to Deep Research prompts.
When Uploaded PDFs Aren't Cited
Deep Research lets you attach PDFs and Google Docs as sources. Sometimes the final report cites none of them. This is also a behavior you can usually correct.
Check three things in order:
File size. Empirically, files over about 20 MB get cited noticeably less often. Heavy scanned PDFs benefit from being converted to plain text first, or split into chapter-sized chunks before upload.
Prompt-level priority. Without an explicit instruction, Gemini treats uploaded files as "supplementary." Add a line like "follow the structure and terminology of the uploaded PDF; treat web sources as supporting context only," and citation rates jump.
Text extractability. A scanned PDF without an OCR layer is unreadable to Deep Research, just as it would be to any text-based tool. Run the file through a vision model first — see our practical guide to image recognition with Gemini for a workflow — and then re-upload the text version.
Three Fallbacks When Nothing Else Works
If you have done all of the above and the result still does not meet your needs, you may be hitting Deep Research's design limits. The three moves I reach for next:
Split the topic into two and depth-research each independently. Running "current state and challenges of A" plus "three-year outlook for A" as two separate jobs almost always produces deeper material than one wider job. Stitching them together by hand at the end usually beats letting Deep Research do it for you.
Feed the report into a regular Gemini 2.5 Pro chat for a second pass. Paste the Deep Research output into a normal chat and prompt: "identify the three weakest paragraphs in this report and reinforce them with stronger evidence." This often produces the depth Deep Research itself could not reach.
Switch to a paid plan (Google AI Pro / Ultra) — but for the right reason. The difference between Pro and Ultra is mostly parallel runs and daily quota, not raw depth per run. If you only run Deep Research a few times a month, Free tier is often enough. We compare Pro and Ultra realistically in Six Months with Gemini Deep Research — An Honest Review.
If you want the equivalent behavior on the API side, building a custom research agent with Function Calling plus Grounding is a viable path. Some of the error patterns you'll hit there overlap with common Gemini errors and fixes for 2026.
A side note that surprised me: when I switched my primary research language to English entirely (prompt + plan + final report all in English) and used a final translation pass via a separate Gemini 2.5 Pro chat for the Japanese version, the depth-per-minute went up noticeably. The crawler appears to spend less compute reconciling cross-language sources, leaving more budget for actually reading them. If your work allows it, this two-stage flow is worth experimenting with.
One Step to Try Today
If your last few Deep Research runs felt thin, try adding just three lines to the plan on your next run: "prefer English primary sources," "include three explicit comparison targets," and "end with a paragraph self-critiquing the weakest argument."
Those three lines change the density of the report, the quality of citations, and the sharpness of the conclusion enough that it feels like a different tool. Deep Research rewards the effort you put into the plan more than almost any other AI feature I use. A few extra minutes upstream consistently turn a half-hour wait into something worth the wait.