If you also work with foreign clients, knowing that ChatGPT cites you in Italian tells you nothing about what happens when someone asks the same question in English — and they're often two completely different worlds. A company that exports and only measures in Italian is navigating half the map. Comparing your visibility across the two languages reveals opportunities and gaps you'd otherwise never see.
AI cites you in Italian but not in English. Or the other way around. AI visibility is asymmetric by language, and if you export — even just to a European market — measuring it bilingually isn’t an option: it’s the only way to understand where you’re actually selling.
I’ll tell you right away, because it’s the first thing that comes up when I work with a client who exports: ask ChatGPT “best Verdicchio di Matelica DOC producers” in Italian and you get 6 names. Run the same query in English — “best Verdicchio di Matelica DOC producers” — and of those 6, only 2 remain. The other 4 are different. Same AI, same question, two distinct worlds.
In this series I’ve been showing you how to measure visibility in AI answers. Today I’ll explain why monitoring it only in Italian, if you also sell abroad, gives you a false picture.
What multilingual AI visibility means for a producer that exports
A winery in Camerino (MC) that bottles Verdicchio di Matelica DOC sells 40% in Italy and 60% across the United States, Germany, the UK and Japan. The American sommelier looking for the next wine to put on the list doesn’t open Google in Italian. They open ChatGPT in English and ask “small producers Verdicchio di Matelica organic”. If the winery doesn’t show up there, for that sommelier it doesn’t exist.
The point is that generative AI models (ChatGPT, Claude, Gemini, Perplexity) have training corpora weighted differently by language. English weighs more, but Italian sources on a hyper-local topic like Verdicchio di Matelica often weigh more than their English counterpart — because very few English-language articles about Verdicchio di Matelica exist.
Translated: whoever is cited in Italian is cited because many Italian sources talk about them (Gambero Rosso, Slow Food, Vinous in Italian, regional food-and-wine blogs). Whoever is cited in English is cited because specific English sources exist (Wine Enthusiast, Decanter, Wine Spectator, Jancis Robinson). They are two sets that overlap very little.
Why linguistic asymmetry comes upstream of everything else
In the previous articles of this series I talked to you about how to measure AI share of voice, about competitive benchmarking, about how to track citations in generative engines. All true, all useful — but if you apply those tools in only one language, you’re measuring half of your market and ignoring the other.
In the world of Large Language Model research the mechanism is documented: models build entity-topic associations within the linguistic space in which they saw them. If “Winery X” appears 800 times in Italian articles about Verdicchio and 12 times in English articles, AI in Italian will recall it strongly, in English barely.
From this follows, for your business, a concrete consequence: your editorial presence in English is an asset separate from your presence in Italian. It has to be built, measured and monitored as a chapter of its own. If your PR only does Italian press, in English you’re invisible — even if you’re number one in your sector in Italy.
This ties directly to what I told you in the article on backlinks as a citation proxy: citations are linguistic before they are global. And it also connects to the reasoning about the weight of implicit references: a company mentioned 50 times by English food-and-wine blogs without a direct link weighs more, in the eyes of an AI in English, than one mentioned 500 times only by Italian sources.
If your PR only does Italian press, in English you’re invisible — even if you’re number one in your sector in Italy.
The bilingual test you can run in 25 minutes
Let me explain the basic test, the one I run with every client who exports before starting any GEO activity. All you need is to open ChatGPT, Claude, Gemini and Perplexity in two different tabs — one with prompts in Italian, one with prompts in English.
Do this:
- Choose 5 commercial queries from your sector (for the Camerino winery: “best Verdicchio di Matelica DOC”, “organic Verdicchio di Matelica small producers”, “Verdicchio Matelica winery visit tasting”, “Verdicchio di Matelica for restaurants”, “Verdicchio di Matelica for export”).
- Translate them into English keeping the same intent.
- Run each one on 4 AI engines: ChatGPT, Claude, Gemini, Perplexity. Both in Italian and in English.
- For each answer, note the brands cited in the first 3 names.
- Compare the two tables: which brands appear in Italian, which in English, how many appear in both.
The decision threshold is simple and binary: if fewer than 50% of the brands cited in Italian also appear in English, you have a serious asymmetry. If fewer than 30%, you are essentially absent from one of the two linguistic markets.
Honest note: this test is a first step, it doesn’t replace a benchmark done with professional tools across dozens of queries and with longitudinal tracking. But to understand whether you have a bilingual problem, it’s enough.
Search intent isn’t translated, it’s reformulated.
The test I ran on 10 Verdicchio di Matelica DOC producers
Before writing this article I wanted to see for myself how much the asymmetry weighs. I selected 10 Verdicchio di Matelica DOC producers taken from the consortia and the industry guides, ran 4 queries in Italian and the same 4 in English on ChatGPT, Claude, Gemini and Perplexity. Total: 32 answers per language.
Results on a small sample (indicative test, not a study):
- Of the 10 monitored brands, 6 appeared at least once in the Italian answers.
- Only 3 of the 10 appeared at least once in the English answers.
- Of those 3, only one was cited consistently in both Italian and English (3+ mentions per language).
- Among the English citations, 2 Castelli di Jesi brands also appeared that didn’t come up for Matelica in Italian — a sign that the English AI makes less of a distinction between the two DOCs and lumps them together under “Verdicchio”.
Limits of the test: narrow sample, non-exhaustive queries, a single moment of measurement. The pattern, however, is clear: whoever doesn’t work on their English presence doesn’t show up in AI answers in English, and the overlap between the two worlds is minimal.
The mistakes I see most often when a client starts with AI monitoring
People who get into bilingual AI monitoring almost always trip over the same points. I’ll list them because you’ll recognize them immediately.
- Translating the query word for word. “Migliori cantine Marche” isn’t “best wineries Marche”: an American importer searches for “Marche wine producers” or “Italian white wine producers”. Search intent isn’t translated, it’s reformulated.
- Measuring only on ChatGPT. ChatGPT is the one with the strongest English bias. Perplexity in Italian makes very different choices from Gemini in Italian. If you monitor a single engine you don’t see the asymmetry.
- Treating the English version as “secondary”. For those who export it’s exactly the opposite: if 60% of revenue is export, English is the main market for monitoring.
- Confusing language with country. A German importer often queries the AI in English, not in German. A Japanese one too. The set of languages to monitor for an Italian food&wine company that exports is almost always Italian + English, rarely do you need to add anything else.
What to do concretely to close the bilingual gap
First thing: establish the baseline. Run the 25-minute test I described above, on at least 5 commercial queries from your sector, in Italian and in English, on 4 AI engines. Save the answers in a sheet: this is your starting picture.
Second: figure out where you’re weak. If you’re absent in English, the problem is in the sources — you’re not present in English-language industry publications (international wine magazines, guides for importers and distributors, foreign food-and-wine blogs). If you’re absent in Italian, the problem is usually the opposite: you’ve invested in a multilingual site but your Italian editorial presence is poor.
Third: set up quarterly bilingual monitoring. Daily isn’t necessary. Every 90 days re-run the same 5 queries in the two languages, on the 4 engines, and compare against the baseline. If the overlap rate between Italian and English doesn’t rise, you need to act on the editorial pipeline in the weak language — not on the site.
To consolidate your presence in both languages, also remember to look after your Google Knowledge Graph entry and to work on author recognition as an entity: they are cross-language anchors that help the AI connect the same identity across the two linguistic corpora.
Bilingual AI visibility: the thread not to lose
Measuring visibility in AI answers only in Italian, if you also sell abroad, is like measuring revenue only from the domestic market and ignoring exports. Technically the figure is true, strategically it’s useless. AI visibility is asymmetric by language and must be monitored bilingually, always, for those who export.
In the next articles in this series I’ll talk to you about how to structure a competitive AI share-of-voice dashboard and how to track your citations in AI answers longitudinally, month over month. They’re the two pieces that close the circle of strategic monitoring.
