Not Translation: Five Traps in Local Buying Language

Consider a composite example. A US bike-computer brand ships its German storefront through a translation API. A customer writes in, politely: the page promises 30 days to return, the order email cites 14. Support finds dollars, a stem length in inches, free shipping over $50, a one-year warranty, and a headline reading Fahrradcomputer. German riders type Fahrradtacho.

Local buying language is not your English page in German — it is the units, money, return terms, compliance marks, and category words a local buyer already holds, and a model fills its answer only from slots that match.

AI does not translate your page; it fills slots from it

Retrieval does not render your storefront into German. It takes a local question — what costs what to my door, when does it arrive, how do I send it back — splits it into slots, and hunts your pages for a value for each slot.

Values must be local in form: a euro price with VAT inside; a delivery window, not “fast shipping”; a return window with a named payer. Where the form does not match, the model skips the slot or fills it with the safest true sentence: this brand also sells in the category. The second is not an error, just not a comparison, and comparisons are where orders come from.

The worst errors survive translation intact. English “warranty” becomes Garantie, though German splits the idea: Gewährleistung, the statutory defect liability, and Garantie, the voluntary promise. “Return” becomes Rückgabe, while distance selling runs on the Widerrufsrecht. The nouns cross the border; the years and the payer do not — and because it reads fine, the wrong number gets repeated.

Consistency finishes the job: where your page disagrees with local media and listings, the model reads another market, not carelessness. Settle market order first with six variables for market entry.

Five traps, one table

All five are one defect: an empty slot on one sheet. The first three are money and promises — wrong, and the answer computes the wrong thing. The last two decide whether the answer names you at all.

Trap	Machine translation writes	The local slot needs	What the gap costs
Units and sizing	inches, pounds, fluid ounces	centimeters, kilograms, local size chart	Spec dropped, no conversion
Payment and checkout	credit card, priced in dollars	local payment methods, tax-inclusive price	“Import fees may apply”
Returns and warranty	30-day return, 1-year warranty	withdrawal window, shipping payer, warranty type	Wrong return answer
Compliance and claims	“best-selling,” “eco-friendly”	marks, registration numbers, permitted wording	Not cited, or not listable
Slang and naming	the dictionary’s first entry	local category word, real search terms	No query matches

Rows one to four close with a checklist; row five does not: the dictionary’s first entry is grammatical, invisible to search, and carries no search volume — a vocabulary problem; see regional synonyms.

Fix landed price and the return split first: a wrong number gets quoted as fact, a missing centimeter value skipped.

Example: one bike computer, two first paragraphs

Example — brand, specs, and prices are invented.

Before: “The C bike computer uses high-precision GPS, long battery life, and durable waterproofing. Free worldwide shipping and 30-day no-reason returns.” A model extracts a category and a claim — no price, no return rule it would repeat.

After: “Der Tacho X2 von C richtet sich an Rennrad- und Pendlerfahrer: 2,6-Zoll-Display, 68 g, GPS-Aufzeichnung mit etwa 20 Stunden Laufzeit. 129 € inkl. 19 % MwSt., Lieferung in 2–3 Werktagen, versandkostenfrei ab 49 €. 14 Tage Widerrufsrecht, Rücksendekosten übernehmen wir; 2 Jahre Gewährleistung. Schwäche: In Tunneln driftet die Höhenmessung.”

Nothing there is a style upgrade: five slots got real values, and one weakness was added on purpose — the sentence an assistant reaches for on downsides, which a machine-translated page rarely has. It also keeps “30-day return” away from the statutory window: two return promises cannot both be true.

Fill the slots, then argue about prose

The order cannot be reversed: a native writer hired before the slots exist produces empty sentences that read beautifully.

Step	Action	Owner	Passes when
1 Slot sheet	Twelve values buyers check, from competitors, listings, tickets	Local ops, content	One value per slot
2 Units and money	Metric units, tax-inclusive price, shipping threshold, delivery window	Content, pricing	Matches checkout
3 Returns and warranty	Withdrawal days, shipping payer, statutory versus manufacturer	Legal, content	Days, years, payer agree
4 Compliance and naming	Marks, registration numbers, claim wording, category word	Compliance, native reviewer	A local reads it once
5 Extraction test	Ask a model landed price, delivery window, return route, in local language	Content	All three right, sampled twice

Step 5 is the only falsifiable step: one wrong answer sends you back to step 2. Keep units and currency machine-readable, or the model abandons the column, as machine-readable comparison tables explains. Sample repeatedly: one screenshot is an anecdote, which is where an AI visibility baseline starts.

Where this breaks

Sales happen inside a marketplace. The model reads listing fields, reviews, and fulfillment. Fill the seller backend first.

The category is heavily regulated. Food, cosmetics, medical devices: admission and labeling are the gate; language sits behind it.

You have review capacity for one market. Ten machine-translated languages make you a foreign seller in ten places; one market done properly beats ten covered.

English is carrying several countries. UK, US, and Australian pages differ on units, tax, and return windows; one page for all three is wrong in all three.

Proofreading is mistaken for localization. A bilingual intern fixes grammar while every slot stays empty. The page reads local and computes foreign.

The rest of this line sits on the content and semantics hub.

This month, take one market where you only ran machine translation: write the twelve-row slot sheet — landed price with tax, delivery window, withdrawal days, shipping payer, warranty years, local category word — fill real values for one SKU, rewrite that product’s first paragraph. Then ask a model, in the local language, the landed price, the delivery window, the return route. One wrong answer means the page is not finished.