Gemini Deep Research
- Category
- Research
- Type
- commercial
- Source
- Vendor page
- Last reviewed
- 2026-10-17 (what this means)
Citation check incomplete — recheck later.
Summary
Gemini Deep Research is a Google feature, in the Gemini app and as a Gemini API agent, that plans multi-step web research and writes cited reports. In independent evaluations (DeepResearch Bench, ICLR 2026; Mind2Web 2, NeurIPS 2025; six preprints), it ranked at or near the top on report quality and coverage, but completed fewer hard search tasks than the leader, and a larger share of its citation links did not resolve or did not exist. Most tests used 2025 versions (Gemini 2.5 Pro and 3 Pro); one tested the Gemini 3.1 Pro API agent. Several benchmarks used a Gemini model as the AI judge. Overall score 63. Informational only; not a guarantee, endorsement, or certification.
Overall score: 63 · Evidence grade: B · Independent sources: 8 performance, 2 other
Factor breakdown
| Factor | Weight | Score | Note |
|---|---|---|---|
| Completion rate | 30% | Top RACE report-quality score among deep research agents in DeepResearch Bench (48.88) and top quality in DeepResearchEval (8.51/10), but only 18% full success on Mind2Web 2 hard tasks (OpenAI Deep Research 28%), 13.3% hallucinated and 18.5% non-resolving citation URLs (highest of 10 systems, Rao et al.), and 12.9% strict acceptance on expert consulting tasks (Gemini 3.1 Pro). In a preprint robustness test (MisKnow-Agent, July 2026), its reports adopted planted false conclusions in up to 54% of cases when the misleading sources looked authoritative. Judgment call. | |
| Disputes | 20% | No legal or regulatory action specific to Deep Research found. Researchers disclosed indirect prompt-injection issues in the wider Gemini assistant (SafeBreach, Tenable), which Google reports it mitigated; none was shown to affect Deep Research specifically. Judgment call. | |
| Adherence to limits | 25% | Mind2Web 2 authors had to modify prompts because Gemini Deep Research sometimes ignored the instruction to cite sources for all information; 72.94% of cited statements matched their sources (ReportBench). Usage limits and tool scope are documented by the vendor; no independent permission or limit test was found. Judgment call. | |
| Evidence quality | 25% | Eight independent performance sources, two peer-reviewed (ICLR 2026, NeurIPS 2025 D&B), one testing the current Gemini 3.1 Pro agent within the last 12 months. Three of the seven used a Gemini model as the main judge. Judgment call. |
Overall = 0.30×60 + 0.20×65 + 0.25×55 + 0.25×72 = 62.75, shown as 63. See the methodology.
Version scope
This entry covers Gemini Deep Research in the Gemini app and as a Gemini API agent. Five of the seven counted sources tested Gemini 2.5 Pro versions from 2025, one tested Gemini 3 Pro (late 2025), and one tested the Gemini 3.1 Pro API agent (2026). Google’s current versions, Deep Research and Deep Research Max, are built on Gemini 3.1 Pro (April 2026). Deep Research Max has not been tested independently. Tests of earlier versions are counted because the product is continuous and only the underlying model changed (as for OpenAI Deep Research).
AI-judge disclosure
Three counted sources (DeepResearch Bench, DeepResearch Bench II, DeepResearchEval) used a Gemini model as the main AI judge. DeepResearch Bench also used Gemini Deep Research reports as its reference answers. This may favor Gemini’s report-quality scores. In DeepResearchEval, a second judge (GPT-5) gave the same ranking. The 2026 consulting benchmark used human expert graders.
Security and disputes
No lawsuit, regulatory action or security advisory specific to Gemini Deep Research was found. Researchers disclosed indirect prompt-injection issues in the wider Gemini assistant (SafeBreach; Tenable TRA-2025-23), and Google reports it has mitigated them. None was shown to affect Deep Research specifically. In the Gemini app, Deep Research can read connected sources such as Gmail and Drive (vendor documentation), so this class of risk is relevant.
Uncertainties
- The current product uses Gemini 3.1 Pro (April 2026); five of the seven sources tested Gemini 2.5 Pro versions from 2025, and one tested 3 Pro.
- Three sources (DeepResearch Bench, DeepResearch Bench II, DeepResearchEval) used a Gemini model as the main judge; DeepResearch Bench also used Gemini Deep Research reports as its references.
- DeepResearch Bench and DeepResearch Bench II share authors; Rao et al. reuses DeepResearch Bench outputs. Each is counted once as a separate study.
- The ReportBench and DeepResearchEval authors work for companies with competing products; this is disclosed.
- The 3.1 Pro test used the API agent, not the Gemini app.
- Excluded from the count: TTD-DR (Google authors) and DRACO (Perplexity, a competitor that ranked itself first).
- No independent test of usage-limit or permission adherence was found.
- Citation check incomplete — recheck later (0057: four 2025–2026 candidates screened, one counted; an Eye brief communication is paywalled until 2027 and an AMCIS 2026 forecasting paper was read as abstract only; Semantic Scholar still unavailable).
Sources
- DeepResearch Bench (ICLR 2026; arXiv 2506.11763): Gemini-2.5-Pro Deep Research RACE 48.88 (highest of 4 deep research agents), 111.21 effective citations, 81.44% citation accuracy; outputs collected 27–29 April 2025. Gemini-2.5-pro was the judge, and the reference reports were Gemini Deep Research outputs (possible self-preference) — independent performance, accessed 2026-10-17
- Mind2Web 2 (NeurIPS 2025 Datasets and Benchmarks Track): partial completion 0.45, full success 0.18, Pass@3 0.30 on 130 tasks; answers collected April–June 2025; prompt modified because the agent sometimes omitted sources — independent performance, accessed 2026-10-17
- ReportBench (preprint; ByteDance authors, whose company offers a competing Doubao deep research product, not evaluated in this paper): 72.94% cited-statement match rate, 92.21% non-cited factual accuracy, reference precision 0.145 — independent performance, accessed 2026-10-17
- Rao, Wong, Callison-Burch (UPenn, preprint): 11,309 URLs from DeepResearch Bench outputs; 18.5% non-resolving and 13.3% hallucinated URLs, the highest of 10 systems. Reuses April 2025 DeepResearch Bench outputs — independent performance, accessed 2026-10-17
- DeepResearch Bench II (preprint; same USTC team as DeepResearch Bench): Gemini-3-Pro Deep Research 44.60 total (2nd of 8), Gemini-2.5-Pro 41.98; systems available as of November 2025; Gemini-2.5-Pro judge — independent performance, accessed 2026-10-17
- DeepResearchEval (preprint; Infinity Lab, Shanda Group and NTU): highest quality 8.51/10 of 9 systems; 76.62% factual ratio (2nd); collected Aug–Sep 2025; Gemini-2.5-pro primary judge, GPT-5 cross-check kept the ranking — independent performance, accessed 2026-10-17
- Deccan AI Research (preprint, v3 June 2026): Gemini 3.1 Pro deep-research via the Gemini API on 70 expert-authored consulting tasks, graded by human experts; VRS 52.6 (2nd of 3), strict acceptance 12.9%, the most rubric zeros (74) — independent performance, accessed 2026-10-17
- Is Deep Research Reliable? (MisKnow-Agent; Beijing University of Posts and Telecommunications, Shanghai AI Laboratory; preprint, Jul 2026): misleading documents supporting false conclusions were supplied through Gemini Deep Research's file interface with web search on. Reports adopted the false conclusion in 54% of cases with high-authority sources, 46% medium and 8% low; 27% with one misleading document and 50–56% with three to five. Robustness test, not task completion — independent performance, accessed 2026-10-17
- SafeBreach: indirect prompt injection via calendar invites against the Gemini assistant (not Deep Research specifically); disclosed to Google, which reports mitigations — independent other, accessed 2026-10-17
- Tenable TRA-2025-23: prompt injection in Gemini Search Personalization (Medium); Google rolled back the model. Not Deep Research specifically — independent other, accessed 2026-10-17
- Google, 21 April 2026: Deep Research and Deep Research Max built on Gemini 3.1 Pro; vendor benchmark claims not used — vendor or maintainer, accessed 2026-10-17
- Gemini API docs: agent versions, default tools (Google Search, URL Context, Code Execution), MCP and File Search — vendor or maintainer, accessed 2026-10-17
Excluded sources
- TTD-DR (preprint): written by Google researchers (vendor), so not counted.
- DRACO (preprint): published by Perplexity, a competing product that ranked itself first, so not counted.
- Vendor benchmark claims for Deep Research and Deep Research Max are not used as independent evidence.