Short answer. A multilingual prompt set is written market by market by native speakers, not translated from one master list. Each market gets commercial, comparison and problem-first prompts in the phrasing local buyers actually use, a fixed core that is re-run unchanged every period, and a record of who wrote it. Translated sets measure questions nobody asks.
Key takeaways
- A prompt set translated from English measures the question an English speaker would ask, which is rarely the question a local buyer types.
- The fixed core must stay unchanged between periods, because a reworded prompt produces movement that looks like progress and is not.
- Competitor sets differ by market, so a comparison prompt naming the same three rivals everywhere is measuring the wrong contest in most countries.
- Prompt-set authorship is a disclosable fact: ask who wrote each language's set, where they sit, and when it was last revised.
- Set size matters less than stability and coverage; a small frozen set run consistently beats a large one that drifts.
Why can't you translate one prompt set into every language?
Because buyers in different markets name the category differently, and an engine retrieves against the phrasing actually used rather than the concept behind it.
This is the single most common defect in multilingual AI-visibility reporting. A head office writes a prompt set in English, has it translated into eight languages, and reports per-market visibility from the result. The translation is accurate and the measurement is still wrong, because accuracy of meaning is not the property that matters. What matters is whether the string resembles what a real buyer types, and the English phrasing of a commercial question frequently has no natural equivalent elsewhere.
The failure is invisible from the dashboard. Every market returns a number, the numbers differ plausibly, and nothing signals that several of them describe queries nobody fires. The only way to catch it is to have someone who works in the market read the set and say whether a buyer would ask that. Our guide to multilingual GEO across languages covers the content side of the same problem.
What goes into a market's prompt set?
Three kinds of prompt — commercial, comparison and problem-first — written in local phrasing, plus a frozen core that is re-run unchanged each period.
Commercial prompts name the category the way local buyers name it. This is where translation fails most often, because the category name is frequently different rather than translated: an industry may be described by its delivery model in one market and by its output in another.
Comparison prompts pit the brand against the competitors who actually appear in that market. A set naming the same global rivals everywhere misses that the relevant contest in most countries is against regional providers, and a brand can lose every local comparison while the global one looks healthy.
Problem-first prompts describe the pain rather than the product. These catch the queries where a buyer does not yet know the category name, which is where a brand is most often absent and most easily added.
The fixed core
A subset of the set — usually the commercial and comparison prompts — is frozen and re-run verbatim every period. Additions are tracked separately and reported as additions. Without this, period-on-period comparison is meaningless: a rewritten prompt returns a different answer, the number moves, and nobody can say whether the brand gained visibility or the question changed.
How do you keep a set stable without letting it go stale?
Freeze the core, review the whole set on a fixed cadence, and record every change with a date and a reason so movement can always be attributed.
These two requirements pull against each other, which is why most sets fail one of them. A set that never changes slowly stops describing the market as categories get renamed and new competitors appear. A set that changes freely produces numbers that cannot be compared across periods.
The resolution is to separate the two activities. The frozen core carries continuity and is changed rarely and deliberately. A revision log records what changed, when, and why — a competitor acquired, a category term that fell out of use, a new product line. When the core must change, the old and new versions are run together for one period so the series can be stitched rather than broken.
Review cadence is usually quarterly for the core and monthly for additions. The test of whether the discipline is real is simple: ask to see the revision log. A provider that cannot produce one is rewriting prompts between periods, and its trend lines do not mean what they appear to mean. This is one of the questions worth asking before hiring AEO and GEO help.
Who should write each language's prompts?
Someone who works in that market and reviews content in that language, with their authorship recorded alongside the set.
This is an operational question that gets answered organisationally. The common arrangement — a central analyst commissioning translations — produces sets with no market grounding, because the translator is being asked to render a string faithfully rather than to judge whether a buyer would type it. Those are different jobs and the second is the one that matters.
A workable arrangement has a named reviewer per language who sits in the market, writes the initial set from the commercial brief, and signs off revisions. Their authorship is recorded because it is a disclosable fact about the measurement: a buyer evaluating a provider is entitled to ask where the Vietnamese prompts came from, and "a translation vendor" is a materially different answer from "our reviewer in Ho Chi Minh City."
Lifewood authors prompt sets with region-native speakers across 100+ languages, from 40+ delivery centres in 30+ countries, under two independent review passes with timestamped approval records. For how this plays out market by market, see our guide to measuring AI search visibility in China and our generative engine optimisation service.
How do you report results from a multilingual set?
Report per market and per engine, keeping memory and retrieval separate, and never average a region into one figure.
A regional average is the reporting equivalent of a translated prompt: it produces a plausible number that describes no actual market. Asia is the clearest case, because the engine mix differs by country, so an "Asia visibility score" is an average across engines with different user bases and different retrieval behaviour.
Three lines per market and engine carry the useful information. Citation rate says how often the brand is quoted or linked. Entity correctness says whether the brand is described as itself rather than conflated with another company. Share of answer says how much of the response the brand occupies relative to those named alongside it. Memory and retrieval are reported apart, because they move on different clocks and have different fixes — the metric itself is set out in measuring AI visibility by share of answer.
Raw run files are kept. A disputed figure should be inspectable rather than arguable, and a provider unwilling to show the raw output for prompts about your own brand is asking for trust in a number that cannot be checked.