I let an AI agent localize my App Store listing

ReelMeal is my iPhone app that turns TikTok, Instagram and YouTube cooking videos into recipes. In July I handed its App Store listing to an AI agent: Claude Code, with Sonar's MCP server for the keyword data and a small App Store Connect CLI for the writes. Over two weeks it did the whole App Store localization: it researched 35 storefronts, wrote the name, subtitle and keyword field in 29 languages, translated the screenshots and set up rank tracking.

It worked. Then, three weeks later, the same kind of agent pass erased 21 of the top-10 rankings the first one had won, ten of them at #1, in a single day.

This is the whole run, with the numbers from App Store Connect and Sonar, including the part that went wrong.

App Store localization by an AI agent: the short version

• Downloads: first-time downloads went from 19 in the 30 days before the localized version (July 4 – August 2) to 90 in the 30 days after (August 4 – September 2). Outside the US, where I ran no ads, 18 → 73.
• Rankings: of 491 keywords the agent started tracking the day before release, ReelMeal was in the top 10 for 0 on August 3, 30 on August 4 and 47 on August 5.
• Reach: storefronts with at least one download went from 11 to 24, and 29 in the latest 30 days.
• The cost: a second metadata pass, which showed up in search on August 28, removed 21 top-10 rankings in one day, mostly exact phrases from the subtitles the first pass had written.

These are small numbers from one app, and the localized release changed several things at once. The caveats are below, before the lessons.

What the agent worked with

The setup, and a lot of permission:

The starting point was modest. ReelMeal had been live since March 30, 2026, and was getting 1 to 7 first-time downloads a week (App Store Connect, weeks of June 15 – July 27). One caveat on credit: an earlier keyword audit in May covered 11 locales and was done with another ASO tool. The second pass in August started from those notes for some locales.

What the agent did, step by step

1. Research and draft every storefront in one afternoon

On July 21 the agent generated 40 listing drafts in Sonar (the feature has since become the Listing Manager) for 36 storefronts in 29 languages, with 178 target keywords between them. The drafts took 18 minutes (12:13–12:31 UTC). The research behind them shows up in Sonar's API log for that day: 58 keyword-metrics calls through MCP, plus keyword search, suggestions, keyword extraction from the existing listing and an ASO score.

A sample of what it proposed:

The same day it wrote name, subtitle, description and keyword field for all 35 locales into the next version, 1.2, which stayed unsubmitted until August 3.

2. Take the trademarks back out

Look at the German and Italian names above. The agent did what keyword data tells you to do: it used the words people type, and people type "TikTok". App Review guideline 2.3.7 warns against packing metadata with trademarked terms, so on August 2 the agent stripped TikTok, Instagram and YouTube from all 35 keyword fields, including transliterations like "тикток" and "틱톡", refilled the freed characters, wrote brand-free names for German and Italian and rewrote eight subtitles. That's the job a human review pass is for, and the agent did it once it was asked.

For what it's worth, the August 26 pass later put "tiktok", "instagram", "reels" and "youtube" back into three English keyword fields, and that version passed review. The names and subtitles stayed brand-free. One approval isn't a guarantee, so I'd keep trademarks out of anything a user can see.

3. Translate the screenshots

ReelMeal's screenshot set has 8 screens. The agent filled the captions per locale through the Screenshot Studio API (75 translation updates across 31 locales on August 2) and the set was uploaded on August 3 for 35 locales, iPhone 6.9" and iPad 13". It also hit two bugs in my own screenshot tool and one missing feature (exporting PNGs from the server), all fixed the same or next day.

4. Track the keywords before release

On August 3, between 12:27 and 12:42 UTC, the agent added 491 keywords across the 35 storefronts to Sonar's rank tracking. Version 1.2 went live overnight. That timing turned out to be the most useful thing in the whole run: it gave me a "before" reading taken just hours ahead of the release.

What happened when the localized version went live

Version 1.2 was live by 01:02 UTC on August 4. Here is the same set of 491 keywords before and after (Sonar rank tracking, App Store top 10 checked against the stored search results):

The August 3 top-50 and in-results counts come from the first tracking pass, so the top-10 count is the reliable one. About 73 of these keywords were already tracked by other Sonar users, so their history goes back further: ReelMeal wasn't in the top 10 for any of them on any day from July 20 to August 2.

A fair objection: these keywords were picked to match the new metadata, so the old listing was never expected to rank for them. The point isn't that the old listing was bad at them. It's how fast Apple indexed the new phrases: 47 top-10 positions by August 5.

Downloads and impressions, App Store Connect via Sonar:

The 30-day "after" window also contains US ads from August 17 and the next version from August 28. The 13-day row is version 1.2 alone with no ads anywhere: 38 downloads, 2.92 a day against 0.63 a day before.

Weekly, first-time downloads stepped from 1–7 to 19 in the release week and stayed at 13–24 a week through October 11. Outside the US, the share of impressions coming from App Store search rose from 88% to 97%, which is what you'd expect if the new traffic came from people searching in their own language. Comparing July 4 – August 2 with August 4 – September 2, Germany went from 3 downloads to 15 and Japan from 1 to 7.

If you're planning your own App Store localization, keeping your home-market rankings intact covers the part this run didn't test.

The day the agent undid its own work

On August 26 a second agent session, working from the app's repo, rewrote the listing again for version 1.3. It flipped all 35 names to keyword-first ("Video to Recipe AI: ReelMeal") and wrote new subtitles and keyword fields for 13 locales. The US title change is its own story, told in the ReelMeal title case study.

Version 1.3 was resubmitted on August 27 (the date the title post uses), and its new names showed up in search on August 28. That day, 31 keywords that had been in the top 50 disappeared from the results entirely. Twenty-one of them had been in the top 10, and ten had been #1. Some of them:

The pattern is plain once you line them up: most were exact phrases from the version 1.2 names and subtitles, and version 1.3 replaced those fields. "Din digitale kogebog" was the Danish subtitle from the July 21 draft. "Smartes Kochbuch" was the German subtitle from August 2. Across all 675 tracked keywords, the top-10 count fell from 143 to 124 overnight, and it hovered between 118 and 132 until October 11. Only one of the lost keywords came back under the same rank source, and German "smartes kochbuch" reappeared at #5 only after the October 12 switch described below.

Why would an agent overwrite a #1 ranking? The August 26 session's own notes say the Danish and Norwegian storefronts "serve the English fallback — no Danish metadata exists", so it wrote new Danish and Norwegian metadata. But Sonar had ReelMeal at #1 in Denmark for the Danish subtitle all through August, which suggests the Danish metadata was live. My best guess, which I haven't confirmed in App Store Connect: the iTunes Search API still returns ReelMeal's English name for those storefronts, and the agent read that as "not localized" instead of checking the ranking data it had access to.

Version 1.3 also won some. In the US, "video to recipe" entered the top 10 the day after (#10), and its best was #6 on October 9. In Vietnam, "ai nấu ăn" went from #18 to #1 on August 29. In the Netherlands, "recepten opslaan" reached the top 10 by October 10. A rewrite isn't a loss by default. A rewrite that doesn't check what's already ranking is.

What this does and doesn't prove

• Version 1.2 bundled everything. Localized metadata, new screenshots and an app update with its interface in 22 languages shipped together. The download step can't be pinned on metadata alone, though the jump in search impressions points at the listing.
• US numbers include ads. Apple Ads ran in the US from August 17 and brought 12 paid downloads for €33.45 through October 13. The outside-US numbers are ad-free and are the cleaner ones.
• No long baseline. Rank tracking for most of these keywords started the day before release, so there's one "before" reading, not months of them.
• Ranks before October 12 are approximate. Until then Sonar read iOS ranks from the iTunes Search API, which orders results by relevance rather than true App Store position. On October 12 it switched to Apple's true ranking, and the top-10 count moved from 130 to 143 overnight with no metadata change. Compare like with like.
• Small numbers. 19 to 90 downloads a month is a real change for a small app and far too few for statistics. Seasonality can't be checked: the app is about six and a half months old.

Rules I'd give any agent before it touches a listing

If you hand App Store localization to an agent, these are the checks I'd make it run:

1. Track before you ship. Adding keywords to rank tracking the day before release cost nothing and gave the only clean before/after in this run.
2. Read the ranking data before rewriting a field. Before any metadata change, have the agent list every keyword you rank for that depends on a phrase or word it's about to remove. That one check would have flagged the ten #1s. The same rule would have caught the "cooking" token in the US title change.
3. Verify "missing" localizations in App Store Connect. Public search APIs don't always show what a storefront's users see.
4. Keep a human pass for trademarks in names and subtitles. Agents optimize for what people type. Guideline 2.3.7 is about what Apple allows.
5. Let the agent do the breadth, keep the judgment. 35 storefronts in a day is the part an agent is good at. Deciding what not to change is still the expensive part.

The keyword field guide covers the formatting rules the agent followed for each locale.

Questions

Can an AI agent localize an App Store listing?

Yes, if it has keyword data for each storefront and a way to write to App Store Connect. Here an agent researched 35 storefronts, drafted listings in 29 languages and wrote them to App Store Connect the same day. It still needed a human pass to remove trademarked terms before submission.

Does localizing App Store metadata increase downloads?

For ReelMeal it did: first-time downloads outside the US went from 18 to 73 in the 30 days after the localized version shipped, and storefronts with any download went from 11 to 24. The same release also localized the app's interface, so metadata alone can't take all the credit.

How fast does Apple rank new localized keywords?

Fast. Of 491 tracked keywords, ReelMeal went from 0 in the top 10 the day before release to 30 the day after and 47 the day after that. Losses are just as fast: a later rewrite removed 21 top-10 rankings the day it went live.

What can go wrong when an AI agent edits App Store metadata?

Two things went wrong here. Its first drafts used trademarked platform names, and its second pass replaced subtitles that ranked #1, apparently because a public API made those storefronts look unlocalized. Both are fixable with a checklist: a trademark pass and a ranked-keyword diff before every submission.

Want to run the same workflow on your app? Connect Sonar to Claude Code or Cursor for live keyword data in every storefront, and track your rankings before your next release. Start free trial.