How to Use Gemini for Translation: Long Documents, Limits, and Review

Gemini is a good translator for passages, emails, and sections of a document, and it can read a long PDF in one go. It is a poor fit as the only tool for translating a whole book or a formatted document, because what limits a long translation is not how much Gemini can read. It is how much it can write back in one response, and what form that response takes: text in a chat, not a rebuilt PDF, EPUB, or DOCX.

So use Gemini when the hard part is language judgment on a manageable amount of text. Plan for chunking, a fixed glossary, and a completeness check whenever the document is longer than one response. When you need the translated file itself, use a file-level workflow.

This guide covers the Gemini app (gemini.google.com) and the Gemini API, including Google AI Studio. All limits below come from Google's own documentation, checked on 2026-10-11. Google changes them often, so check the linked pages before you plan around a number.

What Gemini Accepts, by Surface

Sources: Upload and analyze files in Gemini Apps, Gemini Apps limits and upgrades, Gemini API document understanding, and the Gemini 3 Flash model page.

Three details in those pages matter more for translation than the headline context window:

• The free app context is small. Google lists 32k tokens for the Gemini app without an AI plan. Its own rule of thumb is that 100 tokens is about 60–80 English words, so a 32k window holds roughly 19,000–26,000 English words. That has to cover your instructions, the source, and the translation together. That is a long chapter, not a book.
• Only PDFs get visual understanding. Google says Gemini uses native vision on PDFs and can understand text, images, diagrams, charts, and tables on the page. Other document types, such as TXT, Markdown, and HTML, are "extracted as pure text", and their formatting context is lost (document understanding). If you upload a DOCX or EPUB, assume Gemini sees the words, not the layout.
• Output is text. The Gemini 3 Flash (preview) model page lists text as its only output type. The app can export a response to Google Docs, or export a table to Sheets (export options). That is useful, but it is still a new document built from chat text, not your original file with the language changed.

The Number That Decides Your Chunk Size Is the Output Limit

Most advice about Gemini and long documents points to the 1-million-token context. For translation, that number is misleading. Reading a 300-page PDF is cheap: Google counts each document page as about 258 tokens, so 300 pages is about 77,400 tokens of input. That fits a Pro plan or the API easily, although it does not fit the free app's 32k window.

Translation is different from summarizing, though. The answer is as long as the source. If the model's output limit is 65,536 tokens, Google's 60–80 words per 100 tokens rule caps one response at roughly 39,000–52,000 English words, and often less once you leave room for headings, notes, or a non-English target language that uses more tokens per word. A long book can easily run past that. The book fits in what Gemini reads, but its translation may not fit in what Gemini writes in one response.

That arithmetic is an inference from Google's published limits, not a measured test. It points to one practical rule: choose chunk size by the output you need back, not by the input Gemini can read. In practice, chunks of one chapter or a few thousand words keep each response well under the ceiling. They also keep the review unit small enough that a missing paragraph is easy to spot.

Long context still helps. Gemini can read the whole document once to build a glossary and a style brief before you translate anything. Google's own long-context guide also warns that when a request has to retrieve many separate items from a large context, the model "does not perform with the same accuracy" as with a single target. A whole-book translation asks for thousands of items at once. Expect better control from many smaller requests than from one giant one.

Gemini Returns Text, Not a Usable File

Before choosing Gemini, decide what "done" means. This table separates jobs where text is the deliverable from jobs where the file is.

The line falls between jobs where reading and judgment are the work and jobs where rebuilding the document is the work. Gemini helps with the first. For the second, translation is only half the task.

A Chunking and Terminology-Drift Checklist

This is a working method, not a benchmark result. It turns the risks Google documents (output ceilings, context limits, weaker multi-item retrieval) into steps you can repeat and audit. Adapt the chunk size to your plan and model.

Before translating

1. Freeze the source. Work from one final version of the file. Note its page or word count.
2. Pick your surface. Free app: plan on one chapter or less per conversation. Pro/Ultra app or API: you can give Gemini the whole document for analysis, but still translate in chunks.
3. Split on natural boundaries. Use chapters or sections, not page breaks, so that sentences and footnotes stay whole. Number the chunks (C01, C02, ...) and record each chunk's first and last sentence.
4. Build the glossary first. Ask Gemini to list recurring names, places, terms, and forms of address with a proposed translation. Have a qualified person approve it. The terminology consistency guide covers what belongs in a glossary.
5. Write a one-paragraph style brief. Audience, register, how to treat units and quotations, and what never to translate.

For every chunk

6. Start each chunk with the same header. Paste the approved glossary and the style brief into every request. Do not rely on the conversation remembering them. A new chat, a free-plan context limit, or a long thread can push early instructions out of view.
7. Translate one numbered chunk per response. Ask Gemini to begin the output with the chunk ID and to end with END C07 (or the matching ID), so a cut-off response is obvious.
8. Check length and boundaries. Compare the paragraph count, then confirm that the last source sentence has a translation. A missing END marker means the response stopped early. Re-request that chunk; do not continue it from the middle.
9. Log new terms. If a chunk introduces a term that is not in the glossary, add it, approve it, and use it in later chunks. Do not let Gemini invent a second translation for it.

After all chunks

10. Search for drift. Search the assembled translation for each glossary term and each source-language name. Every hit should use the approved form. Variant spellings and leftover source-language words are the most common signs of drift.
11. Reconcile completeness. Every chunk ID should be present once, in order, with no duplicates.
12. Review risky passages with a person. Have a qualified reader check names, numbers, dates, negation, and anything legal, medical, or publication-bound.

App versus API, in one line: in the app you do steps 6–8 by hand, and the export-to-Docs button is your assembly step. With the API you can script steps 6–11, but the output is still text that you must place back into a document yourself.

A Prompt Pattern That Survives Chunking

Use the same four-block structure as in our ChatGPT translation guide, with chunk control added:

TASK
Translate chunk C07 from [source language] to [target language].

AUDIENCE AND PURPOSE
[Who will read it and what they need to do with it.]

CONSTRAINTS
- Apply the glossary exactly. If a term is missing, keep the source term in
  [brackets] and list it after the translation as NEW TERM.
- Preserve names, numbers, dates, citations, and paragraph boundaries.
- Do not summarize, skip, or add explanations.
- Start with "C07". End with "END C07".

GLOSSARY
[approved term list]

SOURCE
[chunk text, or "pages 41–58 of the attached PDF"]

Keep review separate from translation. When a chunk is done, open a new request with the source chunk, the translation, the glossary, and a short rubric: omissions, additions, changed numbers or names, glossary violations, and wrong register. Gemini's findings are things to check, not verdicts. Like any model, it can miss errors or flag correct choices.

For fiction, voice drift between chapters is a separate risk from terminology drift. The AI novel translation workflow covers character bibles and voice tests.

Check Privacy Before You Upload

What happens to an uploaded document depends on how you use Gemini:

• Gemini app: Google says a subset of chats is read by human reviewers, and its privacy notice asks you not to enter confidential information you would not want a reviewer to see. Turning off Keep Activity stops future chats from being used to improve Google services, but Google says it still uses chats, with help from human reviewers, to help protect Google, its users, and the public (Gemini Apps Privacy Hub).
• Free API and Google AI Studio: The Gemini API terms class free use, including direct use of Google AI Studio, as Unpaid Services. Google may use that content to improve its products, human reviewers may read it, and the terms say not to submit sensitive, confidential, or personal information. Paid API use is not used to improve Google's products (Gemini API terms).

For contracts, manuscripts under NDA, or client files, settle this before you upload anything. The document translation privacy checklist lists the questions to ask any tool, including ours.

When to Stop Using Gemini as the Main Tool

Gemini is the right tool when you can see and repair every failure: a chapter you will read, a passage you want options for, a glossary you are drafting, a translation you are reviewing. Which model is strongest for your language pair is a separate question. Our best LLM for translation guide explains how to test that rather than trusting a leaderboard.

It is the wrong main tool when the job is a complete file. At that point you would be managing dozens of chunks by hand, re-pasting a glossary, checking end markers, and rebuilding layout yourself. That is the work BookTranslator is built to do. You upload the PDF, EPUB, DOCX, or other supported file, it translates the whole document with context and a consistent glossary, and you get a translated file back, plus a bilingual version where the format supports it. For scanned PDFs, OCR mode rebuilds a clean translated PDF from the recognized content rather than copying the original page layout exactly. You see the price before checkout, so you can upload the file and decide from there. Whichever tool you use, keep the review steps above for anything important.