To localize product data for a multilingual ecommerce chatbot, translate what shoppers read while keeping the catalog records behind those translations unchanged. An English request for a 120 cm desk in “Natural oak” and a Lithuanian request for “Natūralus ąžuolas” should resolve to the same variant ID and SKU when they refer to the same item. The chatbot can adapt the wording and number format without changing the finish, dimensions or configuration.
Delivery and returns need a separate lookup. Use the language to decide how to answer and the destination market to decide which policy applies. If a Lithuanian delivery policy is missing, use a verified policy for that same market only if its wording is approved for use. Otherwise, explain what cannot be confirmed and offer a handoff. Never borrow an English-market promise just because it is available.
1. Give each product field one job
Start with a field map rather than a translation brief. For each value the chatbot might mention, record where it comes from, what it describes and whether it can change with language or market. Do this for one variable product before applying the map to the rest of the catalog. A desk with two widths and several finishes will expose more problems than a product with no options.
- Identity fields: Keep the model number, SKU and variation ID exactly as supplied by the catalog. A change from English to Lithuanian should not generate a new identifier.
- Product facts: Use the manufacturer’s dimensions and the attributes on the selected variant. Translate the attribute names and their display labels, not the values they identify.
- Display fields: Localize names, care instructions and number formatting. Read the localized text against the source to check that it makes the same claim.
- Operational fields: Obtain delivery and returns terms from the applicable market rules. Use the product or destination details those rules require instead of relying on a description written for another storefront.
Illustrative catalog record: A 120 cm natural-oak-finish desk has variation ID 4812 and SKU DESK-120-NO. The English label is “Natural oak”; the Lithuanian label is “Natūralus ąžuolas.” Both labels point to the same finish on variation 4812. A 140 cm desk with that finish is a different variation, even if both languages display the same finish label for it.
Use the field map to settle disagreements before launch. If a translated title says 140 cm but the selected variation says 120 cm, correct the title rather than asking the bot to decide which number sounds right. If a localized description claims “solid oak” but the catalog confirms only a finish, remove that claim. A practical review question is: if every translated label disappeared, could the store still identify exactly what the shopper selected and which facts it can state about that item?
2. Translate the labels shoppers use to make a choice
Build a small English–Lithuanian glossary from the attributes shoppers actually use to narrow a selection. For a desk, start with “width” and “plotis,” “finish” and “apdaila,” and the names of the available finishes. Keep “Natūralus ąžuolas” as a display label if that is what the product page uses. The chatbot can still recognize a shopper who asks for “natūralaus ąžuolo apdaila,” an inflected form that refers to the same finish.
Give each glossary entry a display label and a short list of expressions the bot may accept as equivalent. Treat broader language differently. A shopper who says “oak” or “ąžuolas” might mean a finish, a veneer or a requirement for solid wood. If the catalog has several plausible matches, ask which attribute matters rather than assigning a meaning to the word. If the shopper explicitly asks whether the desk is solid oak, answer from the material field, not the finish label.
Translate care directions with the same discipline. Illustrative source instruction: “Wipe with a soft, dry cloth” can be rendered as “Valykite minkšta, sausa šluoste.” The point is to preserve both “soft” and “dry.” Do not expand it into a claim about water resistance or recommend a cleaning product the source does not mention. Keep the care text attached to the product or finish it actually covers, especially if neighboring variants have different instructions.
Review the terms in short conversations as well as on a spreadsheet. First, ask for the finish using the exact display label. Then ask in a full sentence with a different word order or grammatical form. Finally, use a broader or ambiguous expression and verify that the bot asks a useful follow-up. Compare its confirmation with the product page and cart label. A fluent translation is not enough if it sends the shopper to the wrong option or uses a term they cannot find on the page.
3. Localize measurements without changing their meaning
Keep measurements as source values with their units and named axes, then format them for the shopper. The number 120 is not enough: on a desk, it could be the width, a package dimension or part of a model name. If the manufacturer identifies 120 cm as the width, carry both the number and the meaning into the answer. “Width: 120 cm” and “Plotis: 120 cm” describe the same fact.
Illustrative dimensions: An English answer could say “Width: 120 cm; depth: 60 cm; height: 74.5 cm.” A Lithuanian answer could say “Plotis: 120 cm; gylis: 60 cm; aukštis: 74,5 cm.” The decimal separator changes, not the height. Naming the axes is more useful than returning three numbers separated by multiplication signs, particularly when the shopper is asking about the space available in one direction.
Separate a request for a particular variant from a physical constraint. “I want the 120 cm desk” points to a product option. “It must fit in a space no wider than 120 cm” sets an upper limit. In the second case, state the listed width and leave room for the shopper to consider any unlisted clearance needs. Do not promise that the desk will fit based on its nominal width alone. If a dimension is missing, say which one is missing rather than filling it from a similar variant or confusing it with the carton size.
Use the same distinction for currency: formatting is a display choice, while the amount and currency must come from the applicable storefront or offer. A language switch is not a currency conversion. In testing, compare the source numbers with the rendered answer, including at least one decimal measurement. Check that the bot handles both “74.5 cm” and “74,5 cm” in shopper input without changing the number, and that it never reformats a model number as though it were a measurement.
4. Resolve both languages to the same sellable variant
The chatbot needs to match a complete combination, not just retrieve a product page with the right words. On a variable product, finding the parent desk and finding the 120 cm natural oak version are different outcomes. Use the width and finish together to locate the catalog variation. Once it is resolved, carry that variation ID forward instead of rebuilding the selection from its translated label at each step.
Illustrative requests: “Show me the 120 cm desk in natural oak” and “Parodykite 120 cm pločio rašomąjį stalą su natūralaus ąžuolo apdaila” should both identify variation 4812, SKU DESK-120-NO, if that combination exists in the example catalog. The bot can confirm “120 cm desk, Natural oak” in one conversation and “120 cm pločio rašomasis stalas, apdaila – natūralus ąžuolas” in the other. Compare the underlying IDs, not the literal confirmation strings.
Decide what to do when one attribute is missing or the combination does not exist. If the shopper specifies natural oak but not width, ask which width they want when more than one is offered. If natural oak is listed for 140 cm but not 120 cm, explain which constraint prevents the match. Offer an alternative by naming what changes: a wider desk in the same finish, or the requested width in a different finish. Do not quietly select either one.
Keep the identity check distinct from the availability check. An existing variant may not currently be available to order, and availability shown earlier in a conversation may need to be checked again when the shopper acts. In a review, distinguish “this combination does not exist,” “this combination exists but is unavailable” and “availability could not be confirmed.” Those answers guide different next steps. The first invites a different combination, the second lets the shopper consider alternatives, and the third calls for verification rather than a substitute presented as a match.
5. Choose delivery and returns rules by market, not language
Use separate inputs for conversation language and shipping destination. A shopper can ask in English about delivery to Lithuania, or in Lithuanian about delivery elsewhere. In either case, the language decides the wording of the reply, not the terms. If the destination has not been established and the answer depends on it, ask for the country first. Request a postcode only when the available rule or delivery check needs one.
Make the scope of each policy explicit in the records the chatbot can consult. Note the market, the delivery method and any conditions needed to apply the rule. For English and Lithuanian terminology, distinguish “dispatch” from “delivery,” which can be rendered as “išsiuntimas” and “pristatymas.” Distinguish “pickup” from “courier delivery,” or “atsiėmimas” from “pristatymas per kurjerį.” Review these terms in context: a dispatch estimate is not an arrival promise, and a collection option is not delivery to the shopper’s address.
Illustrative exchange: A shopper asks in Lithuanian, “Ar pristatote į Kauną?” (“Do you deliver to Kaunas?”). The bot should look for a policy or delivery check that covers the Lithuanian destination and the selected item. A generic “Delivery available” label on an English product page does not establish coverage. If the policy calls for a postcode check, ask for it before confirming service. If the bot has found a general policy but cannot verify its application to the desk, explain that distinction.
Apply the same process to “returns” and “grąžinimas.” Use the policy for the applicable market, and do not transfer a condition from a different market simply because the products share a SKU. Verify the source for each important clause, not just the page title: one paragraph may cover delivery while another covers collection or a separate service. When someone changes the destination mid-conversation, keep the chosen product intact but recheck the delivery and returns answer under the new market.
6. Demonstrate a fallback for a missing Lithuanian policy
A missing Lithuanian page can mean two different things. The store may have a verified policy for Lithuania but no approved Lithuanian wording, or it may have no verified policy for Lithuania at all. Check the market scope before treating the gap as a translation task. Even a perfectly translated English policy is the wrong source if it covers only a different destination market.
- Look for the right market and language. Use the Lithuanian policy for Lithuania if it exists and covers the shopper’s question.
- Check for an approved same-market fallback. If the verified policy for Lithuania is in English, use approved Lithuanian wording where available. If the store permits sharing the English wording instead, label it as English and make clear that the terms do apply to Lithuania.
- Stop where verification stops. If the only available policy covers another market, or the relevant clause cannot be confirmed, state what is missing and offer an available support route. Do not imply that delivery is unavailable merely because the bot cannot verify it.
Illustrative fallback: The shopper has selected the example desk and asks when it will arrive in Lithuania. The Lithuanian policy page is missing, and the only verified delivery terms cover another market. The bot replies: “Negaliu patvirtinti šio stalo pristatymo termino Lietuvoje. Jei norite, galiu nukreipti jus į klientų aptarnavimo komandą.” In English: “I can’t confirm the delivery timeframe for this desk in Lithuania. If you’d like, I can direct you to customer support.” The response makes neither a delivery promise nor a claim that shipping is impossible.
Where a handoff is available, pass along the chosen variant, destination and exact unanswered question so the shopper does not have to repeat the selection. Record internally whether the gap was a missing translation, a missing market policy or a missing clause. Each calls for a different fix: approved wording, a verified market rule or clarification from the policy owner.
7. Keep voice selection, the accepted cart action and payment separate
Selection is the moment the assistant understands which item the shopper wants. In voice, “the one-twenty natural oak desk” may be clear in context, but a short read-back makes the understood width and finish explicit. If the measurement or finish is uncertain, clarify before presenting a cart action. The same approach works in Lithuanian: use the shopper’s language to confirm the choice while retaining the same variant record behind it.
Adding to the cart is a separate action. A spoken preference, a suggested product card or an illustrative cart preview is not evidence that the shopper accepted the action or that the cart updated. After the shopper accepts, check the reported cart result before saying the item was added. If the result cannot be confirmed, say so and help the shopper check the cart rather than repeating the action blindly or treating a preview as the live cart.
Illustrative language-switch test: The shopper chooses the 120 cm Natural oak desk in English and accepts the add-to-cart action. The store reports that one unit of variation 4812 was added. Just before checkout, the shopper asks to continue in Lithuanian. Verify that the cart still contains one unit of variation 4812 with SKU DESK-120-NO. The summary can render the finish as “Natūralus ąžuolas” without looking up a new variation, clearing the cart or adding a second unit.
Payment is a separate outcome again. “Added to your cart” and “ready for checkout” do not mean the shopper has paid. Only describe payment as completed when the checkout or payment system confirms that state. In the same test, check that changing the language does not trigger a payment claim or disguise a pending step as complete. If the shopper also changes the shipping destination, keep the existing cart item and verify the new delivery terms separately. Language continuity is about preserving the selection, not freezing every checkout condition.
8. Run paired conversations and compare the records
Build the first test around an item you can inspect in the catalog. Record the expected variation ID, SKU, finish, width, other displayed dimensions and availability before opening a conversation. Also record which policy should apply at each destination you will test. This gives reviewers a reference independent of what the bot happens to say in either language.
- Match the same desk. Ask for a 120 cm natural oak desk in English and then in Lithuanian in a fresh conversation. Compare the returned SKU, available finish, width and labeled dimensions. Check that the variant record matches even when the wording and decimal separator differ.
- Change the market, then isolate the language. Ask about delivery for each market and verify the actual policy source and its scope. Then ask about the same destination in both languages. The policy should stay the same even though the answer is localized.
- Exercise the missing-policy path. In a test scenario where the Lithuanian policy is absent and no same-market rule is verified, expect a clear limitation or a handoff, not a promise translated from the English-market page.
- Switch just before checkout. After accepting the cart action, continue in the other language and compare the actual line item, quantity and status. Treat a mock preview as illustrative, not as proof that the cart changed.
Follow the paired run with two negative cases: a finish that is not offered at the requested width and an existing variant that is unavailable. Check that the chatbot distinguishes them and does not invent a compatible option. Add an ambiguous material question such as “Is that oak?” to see whether the bot checks material instead of repeating the finish label.
Save the shopper’s wording, expected values, observed values and policy source for each failure. “Wrong label” and “wrong variation ID” need separate fixes. If the label looks right but the ID differs, the test fails: the localized conversation may be steering the shopper toward another item.
9. Measure product accuracy separately from language quality
Define a passing localization in terms you can verify. A reviewer can judge whether a Lithuanian answer sounds natural, but that judgment cannot establish that the same variation went into the cart. Keep language review separate from checks against the catalog, policy and checkout records. For each check, write down the set of cases it applies to so the rate has a meaningful denominator.
- Variant agreement: Of paired requests that should resolve to the same item, count how many return the expected variation ID and SKU in both languages. Agreement on the same wrong variant does not pass.
- Measurement fidelity: Of tested answers that display measurements, count how many preserve the source values, units and dimension labels while applying the expected locale format.
- Policy grounding and fallback: Of delivery and returns cases, count those that use the correct market source or take the expected fallback when the source is missing. Track unsupported promises separately so they cannot disappear in an average.
- Cart continuity: Of accepted cart actions followed by a language switch, count those that retain the intended variation and quantity without duplicate adds or a premature payment claim.
Review handoffs by reason rather than treating them all as failures. A handoff for an unverified delivery commitment can be the correct result. Several handoffs for the same missing Lithuanian clause point to a policy gap worth fixing; repeated questions about the meaning of a finish may point to unclear labels or an attribute the catalog does not expose.
Rerun the paired desk conversation when a variant mapping, glossary entry, policy or checkout flow changes. Give reviewers a simple release decision: block a wrong SKU or unsupported market promise, and route awkward but accurate wording to language review. That keeps a polished answer from passing when it would produce the wrong order, while allowing copy improvements to be made without rewriting the product facts they describe.
Related guides
- A store-specific checklist for evaluating an AI chatbot
- How to write useful alt text for ecommerce product and preview images
