AI Translation Invented a Competitor Price: Why You Need a Second Reviewer
A second AI reviewer, separate from the translator, found an invented price, missing citations and wrong brand claims. Here is the method and how to give corrections.
Machine translation is fast and mostly good. The mistakes that matter are the ones that read well. This post covers a real case from Dyadfit, a CodesSavvy product, and a review method you can copy.
Dyadfit is a pre-launch waitlist and content site for a muscle-preservation companion for people on GLP-1 medicines. The app is not built yet. The site has 16 guides and 3 calculators in English. Five guides and the protein calculator are also in Spanish, German, Brazilian Portuguese and French.
What went wrong
We translated the English guides with AI. Then a second AI agent, separate from the one that translated, compared each translation line by line against the English source.
It found problems that a fluent read would miss:
- An invented competitor price. The translated text stated a price that was not in the English source.
- Dropped sentences and dropped citations.
- Label claims credited to the wrong brand.
All three would pass a quick read. The text was smooth and confident. Only a line-by-line comparison against the source showed the gaps.
Why an independent reviewer works
- A reviewer that is the same agent as the translator tends to agree with its own work.
- A separate reviewer that starts from the English source has nothing to defend.
- Comparing two texts is a narrower job than judging one text, so it is easier to do well.
To be clear about what this is: the review was done by AI agents. It was not a review by native-speaker humans, and we do not claim it was. It reduces one class of error. It does not replace a human reader where the stakes are high.
How we gave the corrections
The first fix for the brand problem was easy to say and easy to misread. Saying "do not name brands wrongly" leaves room for many readings. The fix that worked was specific: name the molecule, tirzepatide, instead of a brand, so the claim attaches to the right thing.
The same lesson applies when you send corrections to several reviewers or agents at once. A one-line rule gets applied in different ways. A before and after pair gets applied the same way each time.
| Vague | Exact |
|---|---|
| "Fix the brand names." | Change the brand name in the label claim to "tirzepatide". |
| "Do not invent numbers." | Delete the sentence that gives a competitor price. The English source has no price. |
| "Keep the citations." | Restore the citation after the second sentence of the paragraph, as in the English source. |
What to check
- Is the reviewer a separate agent from the translator, working from the English source?
- Is the reviewer asked to compare sentence by sentence, not to rate the style?
- Are all numbers, prices, names and citations matched against the source?
- Are corrections written as exact before and after text?
- Does a test scan translated pages for leftover English, and did you add the new strings to it?
- Is it clear to readers how the translation was reviewed?
For the wider build, see the Dyadfit case study. If you want help setting up a review step for your own content, book a call.
Frequently Asked Questions
Need this built?
Book a free 30-minute call. We'll discuss your goals, give you honest advice, and a clear estimate — no obligation.
Ways we can help
MVP Development
Production-ready MVPs in 4-6 weeks. Fixed price, senior engineers.
Learn moreAI App Rescue
Your Lovable, Bolt, or v0 app breaks with real users? We audit, fix the gaps, and hand back code you own.
Learn moreAI Integration
Add AI features to your product — OpenAI, Claude, RAG, semantic search, and copilots.
Learn moreMobile App Development
iOS and Android from one React Native codebase. Fixed price.
Learn moreRelated articles
Translated URLs With English Text: What It Did to Our Indexing
On our own site, blog pages under translated URLs held mostly English text. Here is what the Search Console data showed, what we changed, and what to check on yours.
Read moreProcessHTTP 500 vs 400 Error Handling: How Wrong Status Codes Wake Up Your On-Call
A 500 for a typo in a form field pages someone for nothing. Here is how to pick the right 4xx or 5xx code, and how to alert only on what needs a human.
Read moreProcessWhy Your Date Is Off by One Day: The JavaScript Timezone Bug
A birthday or due date that shows up one day early is almost always a date-only value that got treated as a moment in time. Here is why it happens, where it hides, and how to stop it.
Read more