Search ChatGPT in English for a recommendation in your industry: your name almost certainly doesn't appear, even if you're well positioned in Italian. It's not the site's fault: it's that the AI was trained on a hundred languages with English as the dominant weight, and if your citations exist only in Italian you're statistically invisible to any foreign query. You're losing every international client who uses AI to search for suppliers in Italy — who in your sector might well be the most interesting ones. Building an English-language presence doesn't wipe out the Italian one: the two coexist, and the path starts from zero.
Open ChatGPT and ask in English for the “best horse-riding agriturismo Tuscany”. Look at who gets cited: in all likelihood it won’t be the Maremma farm stay that communicates only in Italian, but the one with a decent English page and a few mentions on English-language travel media. Now ask the same question in Italian, “miglior agriturismo equestre in Maremma”: maybe there the first one shows up. Two worlds, two pools of clients, and a single technical difference upstream: the language of your citations.
Let me explain it concretely, then I’ll tell you how to measure it and what to do if you’re selling horse-riding holidays in Grosseto but would also like to intercept the German, Dutch and American families searching in English.
Why AI models are global and your PR isn’t
Models like ChatGPT, Claude, Gemini and Perplexity aren’t trained on a single country. They’re trained on a corpus that gathers content from more than a hundred languages, with a huge weight for English and decreasing weights for the others. Italian is among the mid-resource languages: it’s there, but it isn’t the dominant one.
In the field of research on multilingual models, Foroutan et al. (2025) studied precisely how models integrate data from different languages during pre-training:
“These models are typically pretrained on data from over 100 high- and mid-resource languages, leveraging the broad availability of multilingual content on the web.”
In other words: these models are pre-trained on data from over 100 high- and mid-resource languages, leveraging the broad availability of multilingual content on the web. The operational consequence for you is that every time a user queries the model in English, the model draws from a pool where Italian content exists, but is a tiny fraction compared to the English one. If your brand lives only in the 2-3% Italian slice of that corpus, for a query in English it has extremely low statistical odds of surfacing.
It’s the same principle I told you about when I discussed how models segment text into pieces: see tokenization. If your linguistic presence doesn’t cover the query’s pool, you’re not “translatable” on the fly by the model: you simply don’t exist for that language.
What it means that languages don’t “cannibalize” each other
There’s an important technical point an entrepreneur rarely hears explained. Adding content in English doesn’t “hurt” your visibility in Italian. It’s not a zero-sum game.
Again, Foroutan et al. (2025) observed this:
“First, we find that combining English and multilingual data does not necessarily degrade the in-language performance of either group, provided that languages have a sufficient number of tokens included in the pretraining corpus.”
In plain language: combining English data and multilingual data doesn’t necessarily degrade in-language performance for each language, provided that each language has a sufficient number of tokens in the pre-training corpus. It follows that a bilingual PR strategy, if done with a minimum of critical mass in both languages, makes you citable in both — without the English presence “stealing” anything from the Italian one.
Translated for your horse-riding farm stay: having a solid English section on the site, a few guest posts on English-language equestrian tourism magazines and a couple of mentions on international travel media makes you lose nothing on the Italian side. It opens a second visibility channel, the English-language one, which today you probably have at zero.
Just an automatic translation of the site and nothing more.
The query test on your property
This is a check you can run in ten minutes, without professional tools. It’s meant to let you see it with your own eyes.
I ran this test on ten horse-riding farm-stay properties between Grosseto and Upper Maremma, chosen among those visible on the first page of Google for “agriturismo cavalli Maremma”. The setup is simple: the same query on ChatGPT, Claude and Perplexity, first in Italian then in English.
- IT query: “Quali sono i migliori agriturismi con maneggio in Maremma per una vacanza a cavallo?”
- EN query: “Best horse-riding agriturismo in Tuscany Maremma for a riding holiday?”
The pattern I observed: the properties cited in Italian are seven or eight out of ten, with high overlap across the three engines. In English, the cited properties drop to three or four, and they’re always the ones with three things in common: English pages indexed on the site, at least one mention on English-language tourism media (such as international travel blogs, equestrian tourism portals), and updated profiles on international specialized booking platforms. The other properties, technically excellent, simply don’t surface in English.
I want to be clear: it’s an indicative test on a small sample, not a study. But the pattern is consistent with what Foroutan and colleagues describe at the pre-training level, and I see it repeat every time I compare Italian and English-language presences for the same sector.
A guest post with a concrete angle (“how we manage 30 horses on a 50-hectare Tuscan farm”) is worth ten translated articles.
What an “international citation” really is for an AI model
Here a clarification is needed, because many entrepreneurs think it’s enough to translate the site into English. It isn’t. The citation that carries weight in model training has specific characteristics, which I’ve already described when talking about backlinks as a proxy for citation and about implicit reference weight.
In short, an international citation that moves the needle is:
- A piece of content on an English-language domain (or German, French, Spanish depending on the target market) with recognized editorial authority
- With your brand name written consistently, accompanied by sector context (not a bare “Agriturismo XYZ”, but “Agriturismo XYZ, a horse-riding farm stay in Tuscany Maremma”)
- Repeated across multiple sources, because frequency builds the probability of being selected as the answer
An English page on your own site doesn’t count by itself as a “citation”: it’s self-declaration. You need others, in English, to talk about you. It’s the principle of author entity recognition applied to the brand: the AI builds your property’s identity from the sum of what third parties say about you, not from what you say yourself.
The mistakes I see most often
In the equestrian tourism and nature farm-stay sector — but it applies to any SME with international ambitions — I see these patterns repeat.
Just an automatic translation of the site and nothing more. You have the site translated into English with a plugin, and you think you’ve “covered” the English-speaking market. But Google and AI models see an isolated English page, with no ecosystem of external mentions in English. You stay invisible to EN queries.
Mentions only on vertical Italian portals. Strong presence on Italian farm-stay portals, Italian booking sites, Italian travel guides. Zero presence on English-language equivalents like international travel blogs, English equestrian tourism magazines, global farm-stay portals. The citation map is monolingual.
Brand name written inconsistently. “Agriturismo La Fattoria dei Cavalli”, “La Fattoria dei Cavalli Farm”, “Cavalli Maremma Ranch”: three different versions, often on the same English pages. The model struggles to consolidate the entity. Fix one, in English too, and use it everywhere.
Zero presence on international niche platforms. For the equestrian sector there are specialized portals and directories (equestrian holiday, horse-riding tour). Often the Italian farm stay isn’t there, or it’s there with a poor listing. These are precisely the sources the AI goes to draw from when the query is specific and in English.
What to do concretely over the next three months
An international citation strategy for an SME isn’t a multinational-scale investment. It takes focus, not a huge budget.
- Produce at least two contributions in English per year on English-language industry media: vertical travel blogs, experiential tourism magazines, equestrian or farm tourism portals. A guest post with a concrete angle (“how we manage 30 horses on a 50-hectare Tuscan farm”) is worth ten translated articles.
- Get listings on two or three specialized international directories in your sector. For the horse-riding farm stay that means horseback riding holidays portals. For a wine shop it would mean English wine tourism guides. For a B2B manufacturer, English-language industry trade media.
- Build an “About” page in English with a consistent brand name, a clear geographic location (“Grosseto, Tuscany Maremma, Italy”), and specific competencies. It’s the anchor point the AI will use when it cross-references external English mentions.
- Comment on or contribute to global tech/tourism media where it makes sense: market observations, booking data, trends. You become a citable source in English without having paid a PR agency.
- Open Google Search Console and look at the queries in languages other than Italian that already bring you traffic. If there are any, you’re already receiving English-language interest that you’re not converting into citations.
I’ll leave you with a simple binary threshold: if searching for your brand on Google with the “sites in English” filter you find fewer than five external results (not yours, not social), you’re below the critical mass to appear in AI answers in English. Above twenty solid external results, you start to be a recognizable presence in the English-language pool. This is a first check, of course — the real analysis requires professional tools that cross-reference mentions, context and the authority of the citing domain.
How it connects to the rest of the series
The international citation strategy is one piece of the bigger puzzle of your visibility in AI answers. If the mentions you receive are all Italian, you cover the Italian pool. If you want the English-language pool — German, Spanish, French — you have to build a presence there, with the same principles of authority and consistency I described for the domestic market.
In the next articles of this series on citations you’ll see how academic citations work as an authority signal, how to get into Wikipedia entries in a sustainable way, and how syndication amplifies a mention across multiple sources. These are three levers that work even better when you apply them bilingually: a Wikipedia citation in English counts for global queries, one on Italian Wikipedia for Italian ones, and the two don’t cannibalize each other.
