Google opened Deep Review to the public on October 14, 2026, giving researchers and people preparing reports, proposals or presentations a new way to request feedback. The company’s announcement says users receive two free document reviews per week.
The practical recommendation: start with a non-confidential draft you are authorized to share and whose feedback you can check. For a work document, permission to upload matters more than the free allowance.
What is free, and what costs extra
The Deep Review portal requires Google Account sign-in and lists Gemini 4 Argon on Medium as the free configuration. For additional reviews, users supply their own Gemini API key and pay for the underlying model calls, with no separate Deep Review fee.
Google’s pricing FAQ gives two indicative costs for a 40-page document: about US$3 using Gemini 3.8 Flash on Low, and US$20 using Gemini 4 Argon on Extra High. These are estimates, not fixed prices or a spending cap; actual charges vary by document.
For repeated use, count the time spent checking suggestions as well as the bill. The published estimates do not establish how much time a team will save.
The launch post lists PDF and plain-text inputs, with multi-file support still forthcoming. The portal FAQ also lists Google Docs and Google Slides. Their import mechanics were not tested for this article, so do not assume a native integration or support for reviewing a whole document collection together.
Retention and training are different decisions
According to the current portal, documents are retained for up to one year for display and are not used for model training without permission. Its FAQ also offers an optional permission for the research team to use data to improve Deep Review.
The launch announcement uses an unconditional no-training statement; the portal includes the permission qualification. Check the current consent choices before uploading rather than treating the announcement as a promise covering every choice.
The distinction matters even if you decline research use: storage and model training are separate questions. A client proposal or internal business plan may still be unsuitable for an external service that keeps a copy. If your organization has not approved that service and retention period, use its existing review process.
Choose by the document and the decision
This is a decision framework based on the published terms and research, not a performance ranking. Google advertises uses including legal agreements and financial reports; those intended uses do not establish reliability in each field.
Your situation | Practical starting point |
|---|---|
A non-confidential manuscript, report or presentation you may share | Consider an additional critique, with time reserved to verify the suggestions. |
Client, personnel or other restricted material | Establish permission for this service and its retention terms before uploading. |
A document supporting a legal, financial or scientific decision | Keep a qualified human responsible for acceptance; use feedback to identify questions, not approve the result. |
Several files that must be checked against one another | Do not plan around the promised multi-file feature before it is available. |
What the scientific evidence actually shows
Deep Review grew out of Google’s Paper Assistant Tool, or PAT. In a June research preprint, the authors report that PAT found 89.7% of 29 known equation or proof errors in a selected set of 26 mathematics and computer-science papers. This was an author-run evaluation of an earlier system, not a test of today’s public Argon configuration.
That result measures how many selected errors were found—not the probability that any suggested correction is right. It does not establish a false-alarm rate for ordinary business documents, and a report with no flagged issue cannot certify that a document is correct.
The paper also describes conference pilots in which feedback went to authors separately from formal peer review. That is a useful model for everyday use: another source of questions during drafting, with final judgment left to the responsible reader.
Make the first review useful
Choose a draft you are allowed to submit. Decide what you want checked—for example, whether a proposal’s conclusions follow from its evidence.
Use the portal’s optional context field to specify the sections or questions that matter. This guides the request; it does not guarantee complete coverage.
Check each proposed correction against the original document and supporting sources before accepting it. Record unresolved points and whether checking the feedback was worth the effort.
For teams considering a custom process across a collection of documents, our guide to Claude’s document-review workflows addresses that different build-and-coordinate decision.
Methodology: AI-assisted reporting and analysis based on Google’s launch material, its public portal and the PAT preprint. Availability statements, pricing, data-handling text and links were rechecked on October 14, 2026. No document was uploaded and no review or benchmark was run.