Replying to 11 Languages of Google Play Reviews With Gemini Without Sounding Like a Bot

The per-language operations table

Here's what I keep in front of me. You can also hand this to Gemini as context.

The rule of thumb: the language of the review body wins over the device language. Even if someone writes in English, if their device language is something else, replying in their native language lands better.

Pin the glossary in JSON and inject it on every call

The most reliable way to stop vocabulary drift is to move the vocabulary out of your memory and into a machine-readable file. I keep exactly one glossary.json and paste it into every Gemini call.

{
  "ad_hide_option": {
    "ja": "広告の非表示オプション",
    "en": "option to hide ads",
    "ko": "광고 비표시 옵션",
    "pl": "opcję ukrycia reklam",
    "th": "ตัวเลือกซ่อนโฆษณา"
  },
  "forbidden": ["ad-free forever", "no ads at all", "ads disappear completely"]
}

The important part is that forbidden lives in the same file. Under AdMob policy, promising "zero ads" where you should say "option to hide ads" becomes very awkward to walk back. So I pass the banned phrases alongside the glossary — and then check for them again after generation.

import { GoogleGenerativeAI } from "@google/generative-ai";
import fs from "node:fs";
 
const glossary = JSON.parse(fs.readFileSync("glossary.json", "utf8"));
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
 
// Generate two drafts with different nuance, with the glossary pinned
async function draftReplies(reviewBody, lang) {
  const model = genAI.getGenerativeModel({ model: "gemini-2.5-flash" });
 
  const prompt = [
    "You help an indie developer draft Google Play review replies.",
    "A human edits the final text. Avoid flattened tone and repeated boilerplate.",
    `Write the reply in ${lang}.`,
    "Use these translations verbatim, character for character:",
    JSON.stringify(glossary.ad_hide_option, null, 2),
    `These phrases are forbidden: ${glossary.forbidden.join(" / ")}`,
    "",
    "[Review body]",
    reviewBody,
    "",
    "Output two reply drafts with slightly different nuance, separated by ---.",
  ].join("\n");
 
  const result = await model.generateContent(prompt);
  const drafts = result.response.text().split("---").map((s) => s.trim());
 
  // Re-check after generation. The prompt instruction alone leaks.
  for (const draft of drafts) {
    const hit = glossary.forbidden.find((w) => draft.includes(w));
    if (hit) throw new Error(`Forbidden phrase leaked into a draft: ${hit}`);
  }
 
  return drafts;
}

Why check twice? Because when the review body itself says "just make it ad-free," the model tends to echo that phrasing straight back. Across the 71 replies, this post-generation check caught 3 drafts — about 4%. That sounds small until you remember these replies sit pinned at the top of a storefront.

Throwing an error may look heavy-handed. But a bad generation can be regenerated, while a bad reply can only be apologized for. I'd rather fail loud.

The same instinct — pinning terminology through structured output — applies to publishing bilingual content. I wrote that up in Generate Japanese and English in One Structured Call to Stop Term Drift.

The minimum viable prompt

Gemini API or the Gemini web app, either works. Every reply starts from this shape.

You help an indie developer draft Google Play review replies.
A human edits the output, so avoid flattened tone and repeated boilerplate.
 
[Hard rules]
- Reply in the language of the review body
- Use the glossary.json translations verbatim for the ad-hide setting
- Never claim "ad-free forever" or "ads disappear completely" (AdMob policy)
- Open with a warm, specific line; no stiff corporate greetings
 
[Review body]
(paste here)
 
[Reply structure]
1. Gratitude and empathy, with warmth
2. A factual action report (we added X in this update / ad frequency is lower than last week)
3. A forward-looking commitment, concrete if one exists
 
[Output]
Two drafts with slightly different nuance.
Keep proper nouns and hard-rule vocabulary identical between them.

The key is forcing two candidates. With one, you just paste it. With two, a human has to decide which one carries more warmth — and that single decision is what strips out the template feel.

Here's a real case from May 10. The reviewer was a Thai user complaining about weak category classification. The two drafts came back like this:

Draft A:
ขอบคุณมากครับ! สำหรับฟีดแบ็กที่ละเอียดมาก ปัญหาการจัดหมวดหมู่
ที่ท่านระบุเป็นเรื่องที่เราเองก็เห็นด้วยว่ายังต้องปรับปรุง
จะค่อย ๆ ปรับปรุงให้ดีขึ้นนะครับ
 
Draft B:
ขอบคุณมากครับ! ที่แชร์ฟีดแบ็กให้ครับ การจัดหมวดหมู่ของสัตว์
รวมถึงสุนัขและแมวที่ปะปนกันอยู่ตอนนี้
เราตั้งใจจะปรับปรุงในอัปเดตถัดไปครับ

A is a generic "we'll improve gradually." B names the dogs-and-cats mixing and commits to the next update. The review body mentioned dogs and cats, so I sent B with one extra ครับ at the end. About 30 seconds of judgment per reply.

I also tried wiring the whole flow into an automated pipeline, and stumbled badly on response latency. That story lives in Automating App Store and Google Play Review Replies with Gemini API — The 8-Second Rule I Discovered the Hard Way.

A dedicated structure for ad complaints

Roughly 40% of reviews are ad-related. If you run AdMob as an indie developer, this is unavoidable, and a weak reply here drags the whole rating down. My template:

[Thanks + empathy] → [We added the "option to hide ads" in this update] → [Where to find it in settings] → [Ongoing commitment]

A real English example:

Sorry the ads were too much. We just added an option to hide ads
in the latest update — you can find it in the app settings.
Thanks for the heads up!

Deriving Russian, Persian, and Thai versions of this template is exactly the kind of task Gemini handles reliably. Whether an absolute claim slipped in is a job for the post-generation check, not your eyes. By review number 30 in a session, your eyes will miss it.

Never use Gemini on hostile or coordinated reviews

The most delicate cases are hostile reviews and suspicious bursts from multiple accounts. Gemini produces something bland and safe; a human writing from emotion produces something worse. I write these by hand, short and factual.

I'm sorry for the trouble this caused. If you can share more detail
about what happened, I'll check whether it can be fixed.
Please reach out through the in-app contact form.

Coordinated patterns — several accounts posting near-identical 1-star text within a few days — get a polite reply and a place on a watch list. What I record in user-review-reply-playbook-ja.md:

• Multiple accounts using the same insult in 1-star posts clustered within a few days
• Oddly identical pinpoint complaints ("dogs and cats in the animal category", "drains the battery")
• Accounts showing a posting history of "(2)" or more
• Long-form complaints that have been pinned at the storefront top for weeks

These stay on the watch list, and if the pattern persists, I consider reporting to Google Play support. This is a human-monitoring problem, not something to automate with Gemini.

Reply priority and country focus

When time is short, this order is fixed:

• Top public storefront reviews (the first thing a new user reads) — check via showAllReviews=true
• Unanswered 1-star reviews (rating recovery and churn prevention)
• Unanswered 5-star reviews (community building) — in each user's native language
• Unanswered 2–4 star reviews (the middle band)

Country priority runs Japan → Taiwan → Italy → Thailand → Korea → Brazil → United States. Taiwan and Thailand carry unusually large user counts, so I always write those replies in Traditional Chinese and Thai. Consistent ครับ in Thai, device-language-based Traditional Chinese for Taiwan — those small choices are where the warmth lives.

A practical constraint: for review bodies in languages I can't read, I run a Gemini pass for an English understanding first, then write the reply back in the original language. Persian, Ukrainian, and Polish all go through that two-step. Two browser tabs — one translating the body, one running the reply prompt.

Cost, for context: each call runs roughly 900–1,200 input tokens including the glossary, and around 300 output tokens for two drafts. On gemini-2.5-flash, all 71 replies stayed in the pocket-change range. My month-over-month numbers are in One Month with Gemini 2.5 Flash: An Indie Developer's Honest Cost and Performance Report.

Hold yourself to the improvement commitments you made

If your reply promises a concrete improvement, file it in progress-and-todo-ja.md or remaining-tasks-ja.md immediately. From the 71 replies in May, my future-update queue gained:

• A search and filter feature (to relieve thousands-of-pages scrolling fatigue)
• Position save and resume inside a category
• Separating dogs and cats out of the generic animal category
• Replacing legacy low-resolution images
• De-duplicating repeated images
• Ad frequency and close-button visibility tuning

After committing to an improvement in a reply, I ask Gemini to "turn this thread into a task entry" and paste the result into remaining-tasks-ja.md. Gemini is good at extracting action items from prose while preserving nuance. Matching what you said in replies to what you actually ship next quarter is non-negotiable — that one extra step closes the loop.

I've come to treat reply text as part of the product, finished with the same care as the app itself. Gemini sketches; the final brushstroke stays human. That distance feels right.

Checklist to bake into your operations

If you're starting fresh:

• Cap sessions at 30–40 replies; spread the work across days
• Pre-stock 2–3 warm opener phrases per language so Gemini doesn't have to regenerate them
• Keep fixed vocabulary and forbidden phrases together in one glossary.json, injected on every call
• Re-check for forbidden phrases after generation (roughly 4% slip past the prompt instruction alone)
• Always request "two drafts with slightly different nuance"
• Don't run Gemini on hostile or coordinated-spam reviews; write those by hand, short, factual
• Track every improvement commitment in remaining-tasks-ja.md; Gemini can help do the extraction
• Always rewrite at least one sentence of Gemini's draft to avoid auto-translate paste detection

If you only do one thing, write the glossary.json. The moment the vocabulary is pinned, both the speed and the composure of your replies change. Thanks for reading.