Your Buyer Asks in Russian, Not English.

A construction hardware factory had run an English site for two years and its inquiry flow was steady. Last year they noticed something: buyers from Central Asia and South America wrote stilted English in their emails, yet they placed orders quickly.

A salesperson asked one of them why. The answer was straightforward. They do read English — but first they had asked ChatGPT in their own language, taken the answer, and only then written to the supplier in English.

So the buyer asks in Russian and buys in English. The factory only ever appeared in the second half.

In AI search optimization this is a coverage problem, not a quality problem. In the language the buyer used to ask, you do not exist. An English site solves being understood. An additional language solves being found.

Boundaries first. This piece covers whether to add languages beyond English, which one to pick, and how to avoid building something that cannot work. It is not about translation quality, and it does not promise that adding a language will produce inquiries, because we have no data for that claim and can only describe the mechanism. If ninety percent of your buyers come from English-speaking markets, put this lower on the list. One last limit: GEO makes an existing language version matchable. It cannot manufacture demand in a market that has none.

How AI decides which language to answer in

This section is the hardest part of the piece, because it is written down by Google.

Google's documentation on managing multi-regional and multilingual sites contains one line: Google Search tries to find pages that match the language of the searcher.

What matters more is the layer beneath: the language is determined from the visible body text of the page. The documentation states explicitly that it does not use the lang attribute and does not use the URL.

That resolves a very common confusion. Plenty of sites carry lang="en" in the html tag and /en/ in the URL while the body text is Chinese. For language matching, those pages count as Chinese. In reverse, a page whose body is Russian but whose lang attribute is wrong still counts as Russian.

So in generative engine optimization there is exactly one hard requirement on language: the text actually visible on the page must be in that language. Changing the attribute, changing the path, adding a language switcher — none of it changes the determination.

One related trap is documented as well: Googlebot does not send an Accept-Language header, and much crawling originates from the United States. That means sites relying on automatic language redirects may never have all their language versions crawled. The reliable pattern is an explicit link that lets the visitor switch, not an automatic bounce.

Three tests that decide whether to add a language

To start with AI search optimization, discard one criterion: how many people worldwide speak a given language. That number has nothing to do with your business.

Three usable tests remain.

First, the shape of your inquiry base. Sort the last two years of inquiries by country. At many export factories in Shenzhen and Dongguan the records are scattered across several salespeople's inboxes, so consolidate them into one table first. If a single country exceeds ten percent and you have not one word in its language, that is a clear signal. The exercise needs no tools and takes about an hour.

Second, the language buyers actually ask in. Ask your salespeople directly: is English a second language in those emails, and has a buyer ever mentioned checking an AI tool first? That question is more accurate than any industry report.

Third, the signal AI has already published. OpenAI's Signals data shows the non-English share of ChatGPT users rising from 34% to 52%, with Spanish at 12%, Portuguese at 8% and Arabic at 7%; Central Asian languages such as Uzbek and Kazakh are growing by multiples. That dataset covers consumer plans only, excludes enterprise and education tiers, and cannot be transferred to Doubao or DeepSeek. It still establishes one thing: people asking in languages other than English are no longer a minority.

Cross the three and the answer usually converges on one language. Resist the urge to add several. Take one all the way.

A word on data basis. CSA Research has a frequently quoted pair of figures: 76% prefer browsing in their own language, and 40% will not buy from a site in another language. Forrester puts preference for the native language at roughly two thirds, while noting that executives will accept English and technical buyers demand full localisation. Both were sampled among overseas consumers and buyers, so they are not conclusions about the Chinese market, and they must not be read as buying in your language produces orders.

Four technical requirements that decide whether the work is wasted

From a GEO standpoint, shipping the language version is only the first step. Configure it wrong and the work is wasted.

First, hreflang must point both ways. Google's documentation on localised versions gives three hard rules: every language version must list itself and all the other versions; the annotations must be reciprocal, and one-way annotations are ignored; and the URLs must be fully qualified, including the protocol. The third is the one most often missed, because a relative path in the annotation is the same as no annotation.

Second, drop the automatic redirect. As noted, a retrieval chain carries no language preference, so an automatic bounce keeps sending it back to the default version. An explicit link the visitor chooses is the dependable pattern.

Third, the language switcher must be openable. On many sites it is a dropdown whose options live inside JavaScript, so a retrieval chain may see nothing but a button. Writing the switcher as ordinary links costs almost nothing.

Fourth, each language version needs its own title, description and keywords. Not a machine translation of the English title, but a rewrite around the phrases buyers in that language actually search. This step decides whether the version has any chance of being matched at all.

Terminology review is the bottleneck here, not the writing. Budget for a reviewer before you start, or the version sits half finished for a quarter while everyone waits for the person who knows the trade.

All four belong to the configuration layer of GEO. Doing them produces no immediate effect, and skipping any one of them leaves the language version half broken.

Why machine translation does not hold up

This one has an official qualitative basis in AI search optimization, and it is not our opinion.

Google's guidance on automatically translated content makes two points: the output may not make sense and may be treated as spam, and poor or awkward translation damages how users perceive the site. The guidance even suggests using robots rules to prevent machine-translated pages from being crawled.

Note the wording: may be treated as, not will be demoted. There is no usable official figure for how much machine translation costs you, and the percentages circulating online are not reliable.

For a B2B factory the practical failure is more specific: industry terminology comes out wrong. Words like tolerance, anodising or injection mould get rendered into something that is not the term used in that trade, in that language. A buyer recognises within a line that the page was not written for him.

A workable compromise: copy the structure, have a human pass over the terminology. Reuse the skeleton — section order, table fields — and have someone who knows the trade review the terms and the key sentences. For a two-thousand-word page that is one to two days of work, and it saves more than half the effort of writing from scratch.

How to judge whether it worked

For acceptance in GEO, do not rely on traffic alone. A new language version almost always starts at zero.

Three indicators are worth watching, in descending order of reliability.

First, does your page appear when the key question is asked in that language. Write three to five real questions in the target language, put them to ChatGPT or whichever assistant dominates that market, and check whether your brand or domain appears. This test can be run immediately.

Second, has the language version been crawled. The indexed page count in search console is the most basic signal. A version that shows zero indexed pages three months after launch has a technical problem, not a content problem.

Third, do inquiries carry traces of that language — a greeting in the local language, or a mention of having seen you in a specific assistant. These are weak signals, and for B2B they mean more than a traffic figure.

When none of the three looks good, do not add a second language. One version taken all the way beats five that are half finished.

Three questions that come up

Does an English site cover non-English markets?

Not the asking part. A buyer may ask in Spanish in ChatGPT, get an answer, and then contact you in English. Your English site catches the second half and cannot catch the first. That is why factories with a solid English site still lose deals in non-English markets: not for lack of competitiveness, but for absence from the answer in that language.

Subdirectory or a separate domain for the language version?

From a GEO angle, a subdirectory such as /ru/ costs less and concentrates signal, which is why most B2B sites choose it. A separate domain gives the brand more independence and has to be maintained separately. Either way, the three hreflang rules apply unchanged. This is a technology choice and does not alter the method.

The market is small. Is a dedicated language worth it?

Judge by share of inquiries, not by country size. Some small countries have concentrated purchasing, and one of them can outweigh several large ones. Conversely, a populous country outside your target market is worth nothing. The tests are in section two; do not rank by population.

Closing

Back to the construction hardware factory. They did not launch five languages at once. They built one Russian version: eight flagship product pages, terminology reviewed by a customer in the region, hreflang complete, and a plain-link language switcher in the header.

Four months later a salesperson received an inquiry that opened in Russian. The buyer said he had asked ChatGPT about the load rating of a type of expansion bolt, and their product page had appeared in the answer.

Not a spectacular result, and it proved one thing: in that language, they had begun to exist.

Yihang moved its focus to GEO in 2024, after years on cross-border B2B and Alibaba.com services. Alibaba.com offers something convenient: one back office can host several language versions, and the platform supplies the translation. That convenience leads many sellers to treat multilingual as a switch to be turned on. A company website has no such switch: the version has to be built, the terminology has to be reviewed, hreflang has to be configured. Work down to this layer of AI search optimization and the barrier is not technical. It is whether you are willing to write the content again, properly, for one market.

Data and effect boundaries: statements about Google Search seeking pages that match the searcher's language, language being determined from visible body text rather than the lang attribute or the URL, Googlebot not sending Accept-Language and much crawling originating in the United States, and the three hard hreflang rules all come from Google's documentation on multi-regional and multilingual sites, in its own terms. Those rules describe Google Search behaviour and do not mean other AI systems determine language the same way. The OpenAI Signals language data covers consumer plans only, excludes enterprise and education tiers, and cannot be applied to Chinese AI products. The 76% and 40% from CSA Research and the roughly two thirds from Forrester were sampled among overseas consumers and buyers; they are not conclusions about the Chinese market and imply no causation. May be treated as spam is Google's own qualitative wording, and no usable official figure exists for any quantitative demotion. The construction hardware factory is an illustrative situation and describes no specific client. Every figure cited is a mechanism description or an observational statistic; none of it predicts results for any business. The methods make an existing language version matchable. They promise no volume of inquiries, no ranking, and no commercial outcome.

References

1. Google Search Central: Managing multi-regional and multilingual sites (Search seeks pages matching the searcher's language; language determined from visible content, not the lang attribute or URL; Googlebot does not send Accept-Language and much crawling originates in the US).
2. Google Search Central: Localised versions of your pages (hreflang: list self and all versions, reciprocal annotations, fully qualified URLs).
3. OpenAI: Signals data (non-English share of ChatGPT users rising from 34% to 52%; Spanish 12%, Portuguese 8%, Arabic 7%; Central Asian languages growing by multiples; consumer plans only).
4. CSA Research: Can't Read, Won't Buy (76% prefer their own language, 40% will not buy from other-language sites; overseas consumer sample).
5. Forrester (roughly two thirds prefer their own language; executives accept English, technical buyers require full localisation).
6. Google Search Central: guidance on automatically translated content (output may not make sense and may be treated as spam; robots rules suggested to prevent crawling machine-translated pages; qualitative wording, no quantitative figure).