One business. Three questions. Three different answers.
Picture a dental practice in Antwerp. Good practice, good website, ranks respectably. Three people go looking for a dentist on the same afternoon:
- A Flemish local asks ChatGPT: "beste tandarts in Antwerpen"
- A French-speaking colleague from Brussels asks: "meilleur dentiste à Anvers"
- An expat asks: "best English-speaking dentist in Antwerp"
Same city. Same practice. Three completely different shortlists — and there's no rule saying our practice appears on more than one of them. It's entirely possible to be named in the English answer and be invisible in Dutch, in a Flemish city, to Flemish patients.
Most business owners assume search visibility is one thing that either works or doesn't. In AI search, it isn't one thing. It's one thing per language, and the versions don't share.
AI citation is binary and per-language. You are cited or you aren't — and that decision is made separately for each language, because the system answers a Dutch question from Dutch-language sources. Your English visibility does not vote in the Dutch election. There isn't a partial credit.
Why visibility doesn't travel between languages
In traditional search you had a soft version of this problem. Your English page might rank badly for Dutch queries, but it existed, it was indexed, it could pick up something.
AI retrieval is harsher. When someone asks in French, the system reaches for French-language sources to build the answer. If your site exists only in Dutch, you are not a candidate for that answer — not ranked low, not a candidate. There's no page two to be on. There's a shortlist of three names, assembled from sources in the language of the question.
This is why the usual reassurance — "but Google can translate, surely the AI can too" — doesn't hold. It can. It largely prefers not to. Research through 2025 and 2026 consistently finds that content written natively in a language outperforms translated English for queries in that language, because a source already written in the answer's language is a safer thing to quote than one the system has to translate on the fly and hope it got right.
Remember what these systems are optimising for: not the best answer, the defensible answer.
Why Belgium is where this breaks worst
Here's the part that should concern anyone selling in this country specifically.
AI retrieval systems carry a built-in assumption: that a language and a market are the same thing. Ask in French, you must want France. Ask in Dutch, you must want the Netherlands. For most of the world this is a rough approximation that mostly works.
Belgium is where it falls over — and this isn't speculation. Search Engine Land's May 2026 analysis of multilingual regions found that Belgian users get conflated French and Dutch jurisdictional defaults, inside a country whose regions operate under genuinely different legal and linguistic rules. Multilingual regions, the analysis argued, are "the canary" — the place where the architectural flaw becomes visible, because two languages share one geography and the system has no clean way to separate jurisdiction from meaning.
Translated into business terms:
- Your French page may be read as targeting France, not Wallonia or Brussels
- Your Dutch page may be read as targeting the Netherlands, not Flanders
- A Brussels query in either language may be resolved against the wrong regional context entirely
And here's the sharp edge. Belgian businesses cannot rely on the system inferring the right context. In a monolingual market you can be sloppy about signalling geography and language and get away with it, because the default guess is probably right. In Belgium the default guess is a coin flip. The signals have to be explicit — which means the businesses that state them explicitly gain far more here than the same work would gain them in Amsterdam or Lyon.
Structural difficulty is opportunity if you're the one who handles it. Every competitor in your market faces the same broken defaults, and almost none of them have thought about it for ten seconds. Belgium is a harder market to be found in — which is exactly why being found here is worth more.
Roughly half the web is English. Your customers aren't.
Nearly half of all websites are published exclusively in English. It's the default, it feels international, and it quietly costs you every market where your customers don't actually think in it.
The rule is blunt: for every market where your content doesn't exist in the local language, you are absent from AI-driven search in that market. Not weaker. Absent.
For a Belgian business this cuts more than one way. An Antwerp company with an English-only site isn't just missing the Dutch answers — it's missing them in its own city, from its own neighbours. Meanwhile the market data has moved: Lokalise's 2026 research found 63% of global companies now optimising localised content specifically for AI search visibility, and 45% reporting stronger AI visibility in fully localised markets than in English-only ones.
The larger companies you compete with have worked this out. That's the group already doing it.
Translation is not localisation, and AI can tell
Run your site through a translation plugin and you have a site in two languages. You do not necessarily have visibility in two languages.
The difference matters because translated keywords are not localised keywords. People don't search the dictionary translation of what you call your service — they search what they call it. A "boekhouder" and a "comptable" and an "accountant" are the same job described by three different sets of expectations, and the words people actually type reflect their market's habits rather than a glossary.
The typical failure modes, in order of how often they appear:
- Partial translation — main pages done, blog and service pages left in the original. Every gap is a query you can't be cited for.
- Untranslated metadata — page titles and meta descriptions left in English. One of the most common reasons translated pages fail to perform in their own language.
- Machine-flavoured text — content that reads as machine-generated underperforms, because quality is assessed in each language independently.
- Translated but not localised — correct words, wrong instincts. No local examples, no local terminology, nothing that says "this is a business here."
There's an additional wrinkle specific to smaller language markets. ChatGPT has been observed running background searches in English even when the question was asked in another language. So a business with content in both its local language and English gives the system two routes to find it — which starts to explain the finding in the next section.
The plumbing that fails silently
This section is brief and unglamorous, but it's where multilingual sites usually break — and it breaks quietly, which is the dangerous part.
hreflang is the signal that tells search engines and AI systems that your Dutch page and your French page are intentional language variants of each other rather than duplicates or accidents. Get it right and each version can be served and cited in its own market. Get it wrong and the cluster degrades — sometimes all of it.
The four failures that account for most problems:
| Failure | What happens |
|---|---|
Missing x-default | The system falls back to the wrong language for anyone who doesn't match a defined locale |
| Broken reciprocals | If page A points to page B, B must point back to A. One break and the whole cluster can be ignored |
| Conflicting canonicals | A page marked as the French version and canonicalised to the English one. Pick one |
| Wrong ISO codes | en-uk doesn't exist — it's en-GB. Invalid codes are simply discarded |
None of this produces an error message. Nothing turns red. Your site looks perfect to you, in every language, while the signals that make each version findable are quietly broken. This is precisely the kind of thing that gets missed when a language is bolted onto a site afterwards rather than designed in — and it's why retrofitting is nearly always more expensive and more fragile than building it in from the start.
Being multilingual is itself a signal
Here's the finding that surprised us most, and it reframes the whole investment.
A 2026 study of how AI recommends local businesses — close to 500 prompts across three countries and six cities — looked at which businesses actually got named. When asking about local businesses in two German cities, virtually every recommended business had a website in both the local language and English.
Read that carefully. Not "the multilingual ones did somewhat better." Virtually all of the recommended ones were multilingual.
The likely mechanism: dual-language content helps AI systems connect the same business to multiple ways of asking the same question. It's more evidence about who you are, in more forms, cross-confirming itself. In a market where residents, expats, commuters and tourists all search differently — which describes Antwerp, Brussels and Ghent precisely — that's not a convenience feature. It's a visibility asset.
Our largest demo build, Terra Viva VZW, is a Belgian NGO site in three languages — EN, NL and FR — with proper per-language indexing rather than a translation widget bolted on top. Not because trilingual looks impressive on a portfolio, but because a Belgian organisation that exists in one language is invisible to most of its own country. That's the standard the market actually requires.
What it costs, and what it's protecting
Plainly:
| What | Price | Notes |
|---|---|---|
| Professional Website | €1,499 | Second language version included |
| Each additional language | €149 | Added at build time |
| SEO | €119/month | Includes technical and structural work per language |
| SEO + AI Search | €199/month | Both layers, save €69 |
Now weigh that against what a missing language actually means. It isn't a percentage decline in a channel. In AI answers it's a binary absence from every query asked in that language, in a country where a large share of your potential customers ask in a different language from the one your site is written in.
One language, added at build time, for €149. The same language retrofitted in two years — new URL structure, reworked navigation, hreflang across every page, metadata rebuilt, content localised rather than translated — costs considerably more and tends to arrive with the quiet breakages described above.
This is the least dramatic argument on this page and probably the most persuasive: it is cheap now and expensive later, and the interval in between is time you spent invisible to half your market.
Test it in five minutes, in both languages
Ask ChatGPT for the best business in your category in your city — in Dutch. Then ask the identical question in French. Then in English. If you're not on all three lists, that's not a mystery, it's a gap with a known cause. Send us what you saw and we'll tell you what's fixable and what it costs.
The bottom line
You don't have one AI visibility. You have one per language, and they don't share.
In most countries that's an interesting technicality. In Belgium — two main languages, one geography, systems that assume language equals market and get it demonstrably wrong here — it's the difference between being findable and being absent in your own city.
The businesses appearing on all three of those shortlists in two years aren't necessarily better. They just exist, properly and completely, in more than one language. That's a decision, and it's cheapest the day you make it.
For how the disciplines fit together, read our flagship guide on AEO vs GEO vs SEO. For why AI names one business over another in the first place, see why ChatGPT recommends your competitor. Or just tell us what you're working with.
Quick answers
No, not reliably. AI citation is decided language by language: a Dutch question is generally answered from Dutch-language sources, a French one from French-language sources. Citation is effectively binary and happens separately in each language. A business can perform strongly in English and be entirely absent from the answers its Dutch-speaking or French-speaking customers receive.
Because AI retrieval tends to assume language and market are the same thing, and Belgium breaks that assumption. Dutch, French and German speakers share one country whose regions operate under genuinely different linguistic conventions, and reporting in 2026 found Belgian users receiving conflated French and Dutch defaults as a result. Multilingual regions expose a structural weakness in how these systems separate language from geography — so Belgian businesses can't rely on the system inferring the right context. They have to state it explicitly.
Translation is the starting point, not the finish line. A translated page earns visibility only if it's complete, technically signalled as a language variant, and written the way people in that market actually search. The common failures are partial translation, untranslated titles and meta descriptions, and missing or broken hreflang signals — each creating gaps where AI systems and search engines simply don't see your content for queries in that language.
A second language version is included in our Professional Website package at €1,499, and additional languages are €149 each. Retrofitting a language onto an existing site later is almost always more expensive and more fragile than building multilingual from the start, because URL structure, navigation and hreflang signals all have to be reworked rather than designed in.
The evidence points that way. Research through 2025 and 2026 indicates content written natively in a language consistently outperforms translated English for queries in that language, because the system prefers citing a source already written in the answer's language over translating one itself. Machine-translated content that reads as machine-translated tends to underperform — so the quality of localisation matters, not just its existence.
Start with the language your customers actually buy in — Dutch for most Flemish businesses, French for most Walloon ones. Add the second national language if you serve across the language border or in Brussels, and English if you serve expats, tourists or international clients. Studies of AI-recommended local businesses in bilingual contexts found recommended businesses almost always had sites in more than one language, suggesting multilingual presence functions as a signal rather than merely a convenience.