How to Review Technical Documents with Google Deep Review

Google Deep Review launched publicly on October 14, 2026. It is an agentic system that checks technical documents for logical gaps, mathematical errors, and unsupported claims. Researchers, lawyers, and engineers can use it to stress-test drafts before submission or publication.

The tool evolved from the Paper Assistant Tool (PAT). PAT supplied private pre-submission feedback on more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. Authors reported that it surfaced significant math errors in more than one-third of papers that used it. Fifty-eight percent of ICLR authors ran entirely new experiments after receiving the feedback. Deep Review expands that capability beyond computer science to legal agreements, financial reports, design documents, and presentations. It is powered by Gemini 4 Argon on the free tier.

You can access it at research.google/deepreview. Sign in with a Google account. Every user receives two complimentary reviews per week. Additional reviews require a Gemini API key and incur standard Gemini API charges only. Documents are stored for one year for your reference and are never used to train models.

What Deep Review Checks

Deep Review uses a four-stage agentic pipeline rather than a single model call. First, a segmenter agent splits the document into logical sections that can overlap. Second, it allocates more compute to dense sections such as proofs or experimental methods. Third, specialized agents review each section in parallel while retaining full-document context. Fourth, a synthesis agent consolidates findings, checks cross-section consistency, and runs external grounding searches to reduce hallucinations.

The output lists major issues, minor issues, suggestions, and a summary. Reviews typically take 10 to 20 minutes. The system is strongest at catching concrete errors in logic and mathematics. Higher-level judgments about novelty or significance remain with human readers. For related AI health research, see our guide to the AMIE Lancet primary care study.

Step-by-Step: Run Your First Review

1. Open research.google/deepreview and sign in with your Google account.

2. Choose Upload document or Paste text. Supported formats include PDF, plain text, Google Docs, and Google Slides. Multi-file uploads are planned.

3. Optionally add custom context that describes the intended audience or specific questions you want checked.

4. Click Upload Document for Review. The job appears in the Your documents table. Wait for the status to change to complete.

5. Open the completed review. Examine the major and minor issues first. Cross-check every flagged mathematical claim or citation against the original text.

6. For volumes above two reviews per week, add your Gemini API key in the settings panel. Expected cost for a 40-page document ranges from roughly $3 at low effort with Gemini 3.8 Flash to $20 at extra-high effort with Gemini 4 Argon. You pay only the underlying model charges.

The free tier uses Gemini 4 Argon on medium compute. Results appear only in your private dashboard.

Tips for Better Results

Upload the most complete draft you can. Incomplete sections generate noisier feedback. For mathematical papers, include all proofs and appendices. The adaptive budget assigns more compute to dense material, so clear section headings help the segmenter.

Treat the output as a starting checklist, not a final verdict. Pilot users reported both true positives and occasional false challenges to correct proofs. Always verify flagged items yourself. Pair Deep Review with human co-authors or domain experts for the final judgment.

If you need programmatic access, the Deep Review API is available. Details appear in the Google APIs Explorer. Rate limits and pricing follow the connected Gemini API key.

Limitations and Privacy

Deep Review does not replace peer review or legal counsel. It can miss subtle conceptual errors and occasionally over-flag correct arguments. The PAT preprint (arXiv:2606.28277) reported an 89.7 percent detection rate on a curated set of 29 known errors across 26 papers, compared with 55.2 percent for a single Gemini 3.1 Pro call. Those numbers reflect the earlier PAT system; public Deep Review includes Gemini 4 Argon and further pipeline changes.

Privacy settings are strict by default. Uploaded documents stay on private encrypted servers and are not used for training. You can optionally grant the research team permission to use anonymized data for improvement. Documents are deleted after one year.

Conclusion

Google Deep Review gives researchers and professionals a practical way to surface concrete technical issues before human review. The free tier of two reviews per week is enough for most individual authors. Connect a Gemini API key only when volume or deeper effort levels justify the cost. Combine its structured feedback with your own expertise for the best outcome.

Sources

• Google Research, “Analyzing technical documents with Deep Review,” 14 October 2026. https://research.google/blog/analyzing-technical-documents-with-deep-review/
• Deep Review product page. https://research.google/deepreview/
• Preprint: Towards Automating Scientific Review with Google’s Paper Assistant Tool. https://arxiv.org/abs/2606.28277