GEO-074·Issue 07·Strategy & Market·74 / 132
Translation, Localization, Native Creation: Three Costs, Three Outcomes
Translation makes a page readable; localization makes it citable by local AI. Where each layer stops, what it costs, and when native creation actually pays.
Issue focusAcross Markets: Multilingual, Localization, and Regional Semantics
A US kitchen-appliance brand ships its German store the way most teams do: the English page goes to a translation vendor, comes back clean, goes live under /de/. Four months on, German sessions are real and orders are not. A shopper asks an assistant “Kaffeemaschine fürs Büro, entkalkbar, mit Thermoblock” — an office machine, descalable — and the answer names three brands sold in German retail. Theirs is missing. The German is correct. The page never states the descaling program, the brew group, the 220–240 V rating, or a VAT-inclusive price; the US source never had to.
Translation makes a page readable to a German buyer; localization makes it retrievable by a German answer; native creation makes it preferred — and no change of translation vendor moves a page up a floor.
Three layers, three production lines
The expensive mistake is treating the three as grades of one output. They take different inputs and pass different tests.
Translation changes language and nothing else: same structure, same claims, same evidence. It is cheap, repeatable, a fair test of whether a market asks the question.
Localization adds the market’s constraints: units, voltage, tax-inclusive pricing, certification, warranty, return window, and the fields buyers filter on. The structure survives; the evidence is re-sourced.
Native creation rebuilds topic, comparison set, author, and citation channels from the local market: a second editorial line that produces local standing, not clearer prose.
The test: if the evidence still belongs to the source market, the page is at translation, even when a native speaker wrote it. “Military-grade” translated gives no one a test to repeat.
A translated page fails at retrieval, not at grammar
Machine translation is not rejected for syntax. It fails earlier, in retrieval, for three fixable reasons.
Vocabulary. Retrieval matches what the buyer typed: regional synonyms, model names. A one-to-one translation of source keywords lacks that layer: a pushchair, a stroller, and a Kinderwagen are one item, and “fürs Büro” rarely exists in the source.
Evidence that cannot be checked. Adjectives survive translation; proof does not. “Industry-leading” returns as an equally confident adjective, still unverifiable, so a question about drawbacks skips it for a page with a value, a condition, and a warranty term.
Cross-language contradiction. If the US page promises a two-year warranty and the German page one, or one price includes VAT and the other does not, the model can cite either and be wrong. A monolingual error pollutes one market; a contradiction pollutes both — a version-control problem, fixed in locale consistency.
Switching vendors changes nothing: the failure is unswapped evidence, not unidiomatic German.
Costs multiply; gains step
| Layer | What changes | Planning cost (Example) | You get | You don’t get |
|---|---|---|---|---|
| Translation | Language | 1× | Comprehension | Shortlist presence |
| Localization | Constraints and evidence | 3–5× | Local retrieval, extractable fields | Local brand status |
| Native creation | Topic, comparisons, author, channels | 8–10× | Local citations, first choice | Product or fulfillment fixes |
The multipliers are planning ranges, not benchmarks. Translation to localization is a step change: readable becomes retrievable. Localization to native creation is marginal: eligible becomes preferred. So finish localization and stop, upgrading only when two signals co-occur: consistently shortlisted but never first, and local off-site citations near zero. If markets are unranked, run the six variables for GEO market entry first.
Example: one coffee machine, three versions
Example — brand, specs, and price are invented; no real product is described.
Translated: “Leistungsstarke Kaffeemaschine mit One-Touch-Bedienung, ideal für Zuhause und Büro, hochwertige Verarbeitung.” Correct German, no filterable field, nothing to extract.
Localized: the fields a German office buyer screens — a 10-cup glass carafe, removable brew group, automatic descaling, 220–240 V, one tax-inclusive price, a delivery estimate, a 30-day return window, spare parts stocked for years. The prose gets worse and retrieval gets better: every field has one value, and page, checkout, and feed agree. Making the list quotable is the answer unit’s job.
Native: the topic becomes whether a small office needs a thermal carafe; the comparison set is three machines sold in Germany; the author is a named local reviewer. It competes for first recommendation, not legibility.
Translation supplies tone, localization supplies fields, native creation supplies a position.
Four steps to set the level
| Step | Do | Owner | Passes when |
|---|---|---|---|
| 1 Set the level | Pick a layer per market from revenue share, decision complexity, and competition | Growth, founder | You can say which markets stop at localization this year, and why |
| 2 Extract fields | Pull 5–8 filter fields from real local queries; map each to evidence | Content, category | Every field has one value on the page |
| 3 Swap evidence | Units, voltage, certification, tax basis, warranty, returns | Content, operations | Page, checkout, and feed agree |
| 4 Test extraction | Ask the model the original question in the local language | Content | Your field appears; if not, the page changes, not the translator |
Step 1 is where budget leaks: teams buy native content for a country that cannot retrieve them. Step 4 is the only falsifiable one; a screenshot is one draw, so sample as you would for an AI visibility baseline. Fields present but ranked low mean localization is done; fields absent mean no native budget rescues the page.
Where this breaks
Selling mostly inside a marketplace with a small catalog: the platform handles language and local fields, so spend on listings, not a native site.
Regulated categories — medical, children’s, batteries: localization is admission, not an upgrade. Skip it and you are out; do it and native creation may be unnecessary.
Low-price commodity SKUs: buyers compare price and delivery; native content earns no premium.
The common misuse is calling native-speaker polish “localization.” Polishing changes language, not evidence, and the same unverifiable claims return.
Machine translation validates demand fine; it should not become the main indexed locale, or contradictions arrive before traffic does.
The wider line sits on the Strategy & Market hub.
This month, take one non-English market and copy ten questions buyers there actually typed, from chat logs, site search, or marketplace Q&A. Ask a model each in that language and list every filter field in the answer. Fix the three your pages are missing — page, checkout, and feed. Change the vendor only after that.