Multilingual GEO: How German, Spanish and Arabic Markets Find You Beyond English

Bottom line first: Most overseas-facing sites only have an English version, yet buyers in German, Spanish, Portuguese and Arabic markets are asking AI their sourcing questions in their own languages. Language priority isn’t about market size, it’s about four variables: where inquiries come from, certification coverage, competitive density and content reuse rate. Doing two languages properly in year one beats spreading yourself across six half-finished ones.

1. An English-only site shuts the door on four markets’ questions

The language mix on an overseas-facing site rarely matches where its inquiries come from. We’ve looked at plenty of factory sites turning over tens of millions a year: 80-plus pages in English, zero in German, zero in Spanish, yet Germany and Mexico together account for a third of their inquiries.

This isn’t a question of whether to go multilingual. It’s a question of what language AI is using to answer your potential customers. When a German buyer asks about tolerance standards in German in a conversational AI, the retrieval layer looks for German pages first. Your English page may be complete, but it rarely makes the shortlist. It isn’t ranking low, it simply isn’t selected.

Here’s roughly how AI entry points and question languages differ by region:

Region	Common AI entry points	Share of questions asked in the local language (observed range)	Language buyers actually use
German-speaking (DE / AT / CH)	ChatGPT, Gemini, Perplexity, Copilot	60%–75%	German
Spanish-speaking (ES / MX / CL / CO)	ChatGPT, Gemini, Copilot	65%–80%	Spanish
Portuguese-speaking (BR / PT)	ChatGPT, Gemini	70%–85%	Portuguese
Middle East (SA / AE / EG)	ChatGPT, Gemini, Copilot	40%–60%	Arabic and English mixed
Russian-speaking	Yandex, ChatGPT	70%–85%	Russian
Japan and Korea	Naver, Yahoo Japan, ChatGPT	60%–75%	Japanese, Korean

These ranges are our estimates from client server logs and the language mix of their inquiries, not authoritative statistics, but the direction is consistent:Outside the Middle East, buyers in most non-English markets ask their questions mainly in the local language. The Middle East is the exception, where Arabic and English are mixed, and English content performs noticeably better there than in other regions.

There’s another mechanism that’s easy to overlook: most AI retrieval layers still depend on traditional search indexes. Whether a local-language query gets answered depends on whether local search engines have a page of yours in that language. An English page is missing from both steps at once.

2. Ranking language priorities: four variables, one total score

Don’t rank languages by market size. Rank them by cost of entry and deal efficiency. Fix four variables and their weights so nobody argues from opinion in meetings.

Variable	Weight	How to pull the data	Scoring criteria (1–5)
Inquiry sources	40%	Tag inquiries from the past 12 months by country and language	Above 20% scores 5; 10%–20% scores 4; 5%–10% scores 3; 2%–5% scores 2; below 2% scores 1
Certification coverage	25%	You already hold CE, UKCA, GS, INMETRO, SASO and NOM certificates	All target-market certificates in hand scores 5; missing one scores 3; missing two or more scores 1
Competitive density	20%	Take 10 sourcing questions in that language and ask 3 models	If almost no Chinese suppliers in the same category show up in the answers, score 5; 1–3 suppliers, score 3; 4 or more, score 1
Content reuse rate	15%	Share of English pages that can be translated directly without rewriting	Above 70% scores 5; 50%–70% scores 4; 30%–50% scores 2; below 30% scores 1

Weighted total = each variable’s score multiplied by its weight, summed, then multiplied by 20 for a maximum of 100. Here’s roughly how the scores came out for a client making precision machined parts:

Language	Inquiry sources	Certification coverage	Competitive density	Content reuse rate	Weighted total	Schedule
German	5	5	3	4	89	First batch
Spanish	4	3	5	4	79	First batch
Arabic	3	3	4	3	64	Second batch
Portuguese	2	3	5	3	60	Second batch
Russian	1	2	3	2	36	Wait and see

So here’s what I recommend:Do only two languages in year one, starting with the two highest scorers, and get the glossary, page templates and acceptance criteria working first. When can you run three at once? Only in one case: all three languages have a content reuse rate above 70%, and you already have someone dedicated to terminology review. Otherwise the wider you spread, the thinner the quality gets in each language, and all six end up half-finished.

This scoring method doesn’t apply in one situation: if your inquiries are heavily concentrated in a single language, above 60%, the marginal return on a second language is low. In that case, go deeper in the one language first — expand your product pages from 20 to 60, say, instead of rushing into a new language. Our site delivery service page lays out the scope and the boundaries, so take a look before deciding where to invest.

3. Not full-site translation: start with the pages that get asked about

Full-site translation is a budget black hole. Translating a 120-page site into 5 languages means long-term maintenance on 480 new pages, most of which will never be asked about. The right approach is to tier your pages first and decide what to translate in the first batch based on whether it will get asked about.

Page type	Translate in first batch?	Share of site (reference)	Why translate it
Category and product pages	Yes	100%	Landing pages for sourcing questions
Spec and selection pages	Yes	100%	Fact density that feeds AI excerpts
FAQ and common questions	Yes	100%	Phrasing closest to real questions
Certification and compliance pages	Yes	100%	The first hurdle in German-speaking markets and the Middle East
Case studies and delivery records	Yes, local cases first	60%–70%	Local projects carry the most credibility
Lead time, MOQ and payment terms pages	Yes	100%	Must-ask items in a sourcing decision
Insight articles and blog	No, second batch	20%–30%	Low reuse, high update frequency
About us and team pages	No, second batch	—	Brand trust layer, can be deferred
Legal pages (privacy, terms, cookies)	Yes	100%	Compliance baseline, can’t be English-only

For a 60-page site in a single language, the hours break down roughly like this:

Work item	Pages	Hours per page	Total hours	Description
Product and spec page translation	22	3	66	Includes terminology replacement and formula checks
FAQ and terms pages	10	2	20	Short sentences, need to be rephrased as local questions
Certification and compliance pages	6	4	24	Certificate numbers need to be verified one by one
Case studies page	8	3	24	Local cases first
Metadata (title and description)	46	0.3	14	Can’t reuse the English version
Glossary building and review	—	—	16	One-time investment, reused later
Total	46	—	164	About 1 person for 1 month

Note the relationship between the last two rows:The glossary is a one-time investment, but it determines the cost of all later languages. Without a glossary, by the third language you’ll have the same spec translated three different ways, and it only gets harder to unify from there.

When can you skip the FAQ? Only in one case: your category questions are highly standardized and your site already has a dedicated selection tool page or spec calculator. Otherwise, the FAQ is the highest-yield page type per page—it maps almost word for word to how buyers phrase their questions.

4. Three technical prerequisites for multilingual

Getting the content right but the technical side wrong means all that work is wasted. Miss any one of these three prerequisites and your multilingual versions will have indexing chaos or compete with each other.

Prerequisite	Common mistakes	The right way	How to check it yourself
URL structure	Using the same address with `?lang=de` parameters to switch languages	Separate subdirectories `/de/` or subdomains `de.`	Open the German page URL in a new window, do nothing, and see if it’s in German
hreflang	Each language page only links to itself	Each language page links to all languages and adds `x-default`	Check the source code to confirm it’s bidirectional; one-way tags are as good as none
sitemap	Cramming all languages into one sitemap	Split files by language, cross-link with hreflang annotations	Check indexed counts by language in Search Console

There’s one more prerequisite that isn’t part of these three, but is just as critical:Language switching can’t rely only on front-end scripts swapping text. If the German body text is swapped in by a script in the browser, some AI crawlers will still read English. How to check: fetch the page once without executing scripts and see if the returned HTML contains the German body text.

Launch order matters too: finish hreflang and per-language sitemaps first, then publish content. Doing it the other way means releasing a batch of pages with no cross-reference tags, then adding them after crawlers have already been through—re-crawling takes an extra 2–4 weeks.

5. Terminology consistency table: the five types of words machine translation gets wrong most often

Can you use machine translation? Yes, but it must go through a glossary. Machine translation without a glossary will get the meaning wrong in industrial categories, and wrong meaning is more dangerous than awkward language—customers will order based on the wrong specs.

Chinese	English	German	Spanish	The most common machine translation mistakes
Accuracy	accuracy / precision	Accuracy	Accuracy	Merging accuracy and precision, losing the distinction between systematic error and repeatability
Tolerance	tolerance	Tolerance	Tolerance	Translated as “error,” shifting the meaning from an allowed range to a defect
Lead time	lead time	Lead time	Lead time	Translated as “delivery date,” losing the sense of a duration
Material and grade	material / grade	Material	material	Keeping only “material,” losing the grade level
Certifications	certification	Certification	Certification	Translating “complies with a standard” as “certified”

The glossary itself has few fields, but every one of them has to be filled in properly:

Field	Example	Required	Purpose
Chinese source text	Lead time	Required	Single anchor
Single English translation	lead time	Required	Cross-language alignment baseline
Single German translation	Lead time	Required	Mandatory replacement after machine translation
Single Spanish translation	Lead time	Required	Mandatory replacement after machine translation
Machine-translatable	No	Required	Anything not machine-translatable must be reviewed by a human
Owner and update date	Cui Lishu / 2026.09	Required	Accountability and traceability when something changes

The process is: run machine translation, then run the glossary replacement, then measure the replacement hit rate.A hit rate below 90% means the glossary doesn’t cover enough. When that happens, expand the glossary first instead of editing sentence by sentence. Editing sentence by sentence costs three to five times more than expanding the glossary, and you’ll have to do it all over again for the next language.

Where do you start with a glossary? Don’t build 500 entries right away. Start with the 60–80 words that appear most often on your site, including parameter names, process names, certification names and units. That usually covers 80% of all term occurrences in your body text.

One exception needs to be made clear: don’t use machine translation plus replacement for legal and compliance text. Privacy policies, terms of service and cookie notices are legal texts, and a mistranslation creates compliance risk. These pages must be translated by a human, with the translator and version information retained.

6. Acceptance criteria: how to compare language versions side by side

Multilingual work is easily glossed over with “it looks fine.” To compare versions side by side, you first have to set the criteria, and don’t use English rankings as a benchmark for other languages, because the index pools and competitive landscapes are completely different.

Metrics	Metric	Observation period	Reference target range
Indexing rate	Indexed pages in that language ÷ published pages	4–8 weeks after launch	No lower than 70%
Crawl coverage	Crawled URLs in that language ÷ published pages	4 weeks after launch	No lower than 60%
Local-language keyword rankings	Pick 20 local-language keywords per language and count how many reach the top 100	8–12 weeks	At least 6
AI mention rate	Ask 3 models 10 local-language questions per language and see whether they mention the brand name	Once a month	At least 2 mentions in the first quarter
Local-language inquiries	Tag inquiry forms by language and count inquiries in that language	3–6 months after launch	At least 1 qualified inquiry in the first quarter

Of these five metrics,the AI mention rate is the first to show which way things are heading It doesn’t depend on rankings — it only checks whether the model mentions you in answers in the local language. If none of your 10 questions get a mention after three months, the problem is usually content coverage, not language quality.

Track this in a horizontal table: one language per row, one metric per column, filled in once a month. That’s far more useful than saying “the German version feels better than the Spanish one,” and you can take it straight into a budget review.

7. Four common mistakes and when not to do this at all

Mistake	What it looks like	Consequence	Fix
Translating only the navigation, not the body copy	The menu is in German, the body text is still English	Nothing in the local language to cite for local-language queries	Translate body copy and metadata together
Switching languages with parameters	`/product/?lang=de`	Tangled indexing, broken hreflang	Use separate URLs
Mismatching language and market	Using Spanish for Brazil	Content doesn’t match local search habits	Use Portuguese for Brazil; use Spanish for Spanish-speaking Latin America
Launching everything at once	Six languages go live at the same time	Terminology and maintenance both spiral out of control	Roll out 1–2 languages per batch

The third mistake is the hardest to spot. Brazil’s official language is Portuguese, so using Spanish for Brazil means serving content to a market that doesn’t search in it. By the same logic, don’t cover the Middle East in Arabic alone — English still has high coverage in business settings in the UAE and Saudi Arabia, so running both language versions is the safer bet.

When should you not go multilingual? Three cases. First, if you only serve the US, UK and Australia, there’s no search volume for local-language versions. Second, if you have fewer than 30 pages of reusable content, translating leaves you with only a dozen or so pages per language — not enough to get indexed. Third, if you have no terminology or review resources and can only push out raw machine translation. The third is the most dangerous: better to skip it entirely than to publish a batch of pages with semantic errors. AI will read those pages and describe your capabilities based on the wrong information, and fixing that later costs far more than not doing it in the first place.

FAQ

Which less common languages give the best return? What order should we do them in?

Rank by language priority score, not by market size. Take the machinery and precision machining clients we’ve worked with: German usually scores highest on the weighted total, followed by Spanish — the German-speaking region has a high share of inquiries and good certification coverage, while Spanish-speaking regions have scarce content supply. But if 60% of your inquiries come from North America, Spanish is worth more than German. So the order isn’t fixed. The first step is always to tag the past 12 months of inquiries by language, then score them against the four variables.

Can we just use machine translation as is?

You can’t ship machine translation as-is, but it works as a first draft. The right approach has three steps: produce a machine translation, force-replace terms from a glossary, then have a human review only the segments where glossary matching failed and anything involving certifications or legal language. Track the replacement hit rate as you go; if it falls below 90%, expand the glossary first. The real risk of publishing with no review isn’t awkward language, it’s mistranslated meaning. Translate “tolerance” as “error” and customers will order against the wrong specs. Privacy policies and terms are the exception. Those must be translated by a human.

Does Arabic really need RTL? Is it technically difficult?

Yes, Arabic requires a right-to-left layout, or both readability and credibility suffer. The technical cost is lower than you’d expect. Mainstream themes and page builders support RTL stylesheets. Check three things: the direction of numbers and units, the column order in tables, and any text embedded in images. The real difficulty isn’t the layout, it’s how you phrase Arabic content for localization. Put Arabic in the second batch and get your process running smoothly with German and Spanish first.

Who writes content for less common languages? Can machine translation plus a student editor work?

Look at it in two layers: terminology and specs need someone who understands the product, and fluency needs a native speaker. A student editor can handle the second layer but not the first. They usually can’t explain the difference between DIN and EN standards, or between “meets the standard” and “is certified.” If your budget is limited, buy native-speaker review first, then translation. Machine translation is only a first draft. Review typically accounts for 20%–30% of total spend per language.

How exactly should hreflang be set up? Is it wrong for a page to point to itself?

Every language page should list all languages and point to itself as well. That’s standard practice, not an error. There are three common mistakes: listing only one direction, where the German page points to the English page but the English page doesn’t point back; leaving out x-default; and using inconsistent language codes, such as de in one place and de-DE in another. After launch, crawl the source once and check each cross-reference for two-way links. It takes twenty minutes and prevents months of indexing chaos.

Should each language get its own subsite?

Most companies expanding overseas don’t need subsites. Subdirectories are enough. They cost the least to maintain, concentrate authority, and make hreflang simplest to configure. They suit sites with under 200 pages maintained by one team. Subdomains or subsites fit two cases: you have a local team on the ground with very different content and pricing, or the main brand needs to be separated from a local brand. Keep in mind that subdomains carry their own authority, so a new subdomain starts from zero and will noticeably slow indexing at first.

Final thoughts

The biggest misunderstanding about multilingual GEO is treating it as a translation project. It isn’t. It’s a project of trade-offs, deciding which pages in which markets are worth answering in the local language and leaving the rest for later.

If you remember only one number, remember this:Two languages done properly in year one beat six half-finished ones. Language count doesn’t compound; consistent terminology and messaging does.

To figure out which two languages to start with, run your last 12 months of inquiry data through the scoring table above. It takes about two hours, and it will save you far more than starting three languages at once. For more on putting multilingual GEO into practice, see the Insights section or just Contact Us get in touch.