Google is making its technical-document review system available worldwide, giving each user two complimentary reviews per week. The public release of Deep Review moves a tool developed for pre-submission research-paper feedback into a broader service for checking complex documents.
The October 14, 2026 announcement describes an agentic workflow that divides a document into sections, allocates computation according to difficulty, runs parallel verification, and assembles a grounded report. Google says higher-volume users can connect a Gemini API key for increased usage limits and programmatic access through the Deep Review API.
For researchers and engineering teams, the practical boundaries matter as much as the launch: input support starts with PDFs and plain text, multi-file support is still forthcoming, and uploaded documents remain stored for one year. Google says those documents are never used to train, fine-tune, or improve its models.
Two Free Reviews, Then a Gemini API Key
Users can start through the Deep Review portal. Google describes the service as accessible worldwide, with two complimentary document reviews per user each week.
That allowance provides a way to assess the tool on a small number of documents before considering a larger workflow. It should not be confused with unlimited free access: Google directs users who need higher usage limits to connect their Gemini API key.
The same key-based route supports programmatic integration through the Deep Review API. Google specifically suggests using it in continuous-integration and drafting workflows, which could make document review a repeatable step rather than an occasional manual upload.
The announcement describes this access model without quoting a per-review price for higher-volume use. It also does not provide a numerical limit for key-connected usage in its launch details. Teams evaluating automation should establish the applicable charges and limits before building around the API, rather than treating âconnect a keyâ as a promise of unlimited capacity.
For an initial evaluation, the free allowance is the clearest offer. Whether the service makes economic sense for repeated manuscript revisions or an engineering-document pipeline will depend on the usage terms beyond that allowance.
Four Stages Divide the Verification Work
Googleâs technical argument is that reviewing a long document requires more than asking a model for feedback in a single prompt. Its announcement identifies several possible failures of that approach: exhausting the modelâs reasoning budget, overlooking dependencies between sections, and accepting nuanced claims without sufficient scrutiny.
Deep Review instead uses a four-stage agentic pipeline:
-
Document segmentation. A segmenter agent divides the input into logical sections, such as methodology, experiments, or an introduction. Sections can overlap, but together they must cover the documentâs content. This gives later agents bounded review tasks rather than leaving one model call to handle every part equally.
-
Adaptive budgeting. The system examines information density and assigns computation accordingly. Googleâs example is a dense mathematical proof receiving more compute than an introductory or related-work section. The practical goal is to spend reasoning effort where the document demands it, instead of applying the same budget throughout.
-
Deep parallel verification. Specialized reasoning agents review individual sections in parallel. Google says each agent retains access to the full document context, allowing it to follow cross-references while concentrating on its assigned section. That distinction matters: simply splitting a paper into isolated chunks could hide contradictions or dependencies elsewhere in the manuscript.
-
Global synthesis and agentic grounding. A synthesis agent combines the section reports, checks consistency across them, and can rerun logical checks. It also performs external grounding checks and tool calls, which Google describes as safeguards against hallucinations.
Google says the public release adds support for Gemini 4 Argon, extending the capabilities of earlier versions of its paper-review tool.
The architecture offers a more specific account of how the system attempts review than a general promise of âAI feedback.â It does not, by itself, establish how reliably the system catches errors. Segmentation, additional reasoning, and external checks are mechanisms for improving a review; they are not proof that every criticism in the resulting report is correct.
For users, the useful output is therefore a set of claims to inspect. A suspected contradiction, unsupported conclusion, or flawed proof step still needs to be checked against the underlying document and evidence.
PDF and Text Support Are Available Now
Deep Review initially accepts PDF and plain-text documents. Google says support for multi-file inputs is coming soon, without giving a launch date in the announcement.
That separates the current product from the broader workflow it may eventually support. A technical proposal with several attachments, or a paper whose evidence sits across separate documents, should not be treated as an already-supported multi-file review package.
Researchers should check that the submitted document contains the context needed to assess its claims. A reviewer cannot reliably evaluate a result if the necessary assumptions or supporting material are absent, regardless of how much computation the review pipeline assigns to it.
Google also describes applications beyond academic papers, including design documents, presentations, and legal or financial material. Those examples describe the kinds of content the company wants Deep Review to examine. They should not be read as a launch-format list or confirmation of native Google Docs and presentation-file integrations.
The immediate planning rule is straightforward: prepare a PDF or plain-text input, and do not build a workflow that depends on promised multi-file support arriving by a particular deadline.
No Model Training Does Not Mean No Storage
Googleâs Deep Review data policy makes two separate commitments. Uploaded documents are never used to train, fine-tune, or improve Google models. They are also stored for one year, which Google says is solely to display them to the user in the Deep Review interface.
The distinction is important for unpublished research, confidential design documents, and commercially sensitive proposals. A commitment not to use a document for model training does not mean the service processes it without retaining a copy.
Researchers and engineering teams should compare that one-year storage period with their institutionâs or employerâs rules before uploading material. This is particularly relevant where an author has permission to draft or read a document but may not have permission to submit it to an external service.
The public serviceâs retention policy also differs from the earlier conference pilots. Google says papers submitted through those programs were deleted promptly after each program concluded. Users familiar with those pilots should not assume the same short-lived storage arrangement applies to the public release.
These are Googleâs stated commitments, not an independent privacy audit. The announcement supports a clear conclusion about training use and the stated storage period, but it should not be expanded into assumptions about other contractual or security guarantees.
Sources
- Deep Reviewresearch.google
- Deep Review portalresearch.google
