Visible in English, Invisible in Your Clients' Language: The AI Search Gap

TL;DR: When a prospective client asks an AI engine to recommend an immigration attorney, the answer it returns depends on the language of the question. The same firm can appear when the question is asked in English and be absent when an equivalent question is asked in the client's first language. For a firm that built part of its practice on serving one language community, this gap tends to open in the place that matters most: the answer the community member sees when they ask in their own language. The fix starts with checking both, not one.

A firm can be named when a client asks ChatGPT for an immigration attorney in English, and absent from the answer when the same client asks in their first language. The two answers are assembled from different material, so they can disagree. If your practice is built partly on serving one language community, the language version is the one your positioning depends on, and it is the one most firms never check.

Why this matters for a firm with a language promise

Many boutique immigration firms make a specific promise: we can represent you in your own language, talk to your family in it, and handle the first hard conversation without a translator in the room. That promise is often the reason a client chooses one firm over a larger, cheaper, English-only competitor.

The promise lives or dies at the moment of search. And increasingly that moment is not a Google results page. It is a question typed into ChatGPT, Perplexity, or Google's AI answers, often in the client's first language, because that is the language they think in when the matter is personal and stressful. If the firm is missing from that answer, the language promise never gets a chance to work.

Why do AI answers change with the language of the question?

Because the model assembles each answer from a different pool of material depending on the language it is asked in.

Large language models are trained on text scraped from the web, and that text is dominated by English. The raw web crawls these models start from are already heavily English - analyses of Common Crawl, the main public source, put English near half of all documents, with every other language splitting the rest. By the time a model is trained, the English share is usually higher still: disclosed figures compiled in a 2024 survey of multilingual model corpora (arXiv:2404.00929) have run from roughly 78 percent English for PaLM to over 92 percent for GPT-3. A language community that is large in the world can still be small in the data the model learned from. When the model answers in that language, it has less to draw on, and the firms it has seen named in that language are fewer.

Retrieval works the same way. When an AI engine answers a current question, it often runs a live search and reads pages before replying. A query in Korean or Portuguese retrieves a different set of pages than the English version of the same query, because the search index returns language-matched results first. If a firm's citable material - directory listings, articles, profile text, third-party mentions - exists mainly in English, it is far more likely to be retrieved by the English question than by the same question asked in another language.

So the two answers are not two views of one ranking. They are two separate assemblies, built from two different piles of evidence. They can name different firms, and often do.

Two patterns we have observed

In reviewing how AI engines answer immigration queries across languages, two patterns recur. Both are described here without firm names, because the point is the mechanism, not any individual firm.

In the first pattern, a business immigration attorney who openly serves a specific non-English community appeared in the English-language answer to a request for an attorney who speaks that community's language. The same request, asked in the community's own language, returned a list of a dozen firms - and that attorney's name was not among them. The firms that did appear in the community-language answer were a mix: a few small firms, and at least one large national practice with no particular language specialty at all.

In a second, mirror-image pattern, a firm serving a different language community was the top name in the English answer and entirely absent from the answer in the community's language. Same firm, same specialty, opposite result depending only on the language of the question.

We would not treat any single result as the whole market. AI answers vary between sessions and shift over time. But the direction is consistent enough to be worth a firm's attention: presence in English is not evidence of presence in the language your clients actually use. For more on how these decisions form in the client's head, see how AI search changes immigration client decisions.

You checked your visibility - in which language?

A first check of AI visibility is real and worth doing. It tells you whether a model can name your firm at all when asked a plain English question about your specialty. That is a genuine signal, and a firm that fails even the English check has a clear starting problem.

But the English check answers a narrow question: does the model know you exist in the language the internet is mostly written in. It says nothing about the question that decides whether your language promise reaches anyone: are you named when a member of your community asks in the language that promise is about. Those are assembled from different material, so one cannot stand in for the other.

The more useful check is the same question in both languages, side by side. The gap between the two answers is the part of your visibility that your single English check could not see.

What opens, and what closes, the language gap

A few factors tend to widen the gap. A practice whose entire web footprint - site copy, directory profiles, articles, third-party mentions - is in English gives the community-language query almost nothing to retrieve. A new or recently relaunched firm has a thin footprint in any language, and the thinnest part is usually the non-English side. A reliance on the attorney's bio alone, with no material written in or about the community language, leaves the model with no language-matched evidence to surface.

A few factors tend to narrow it. Citable material that actually exists in the community language gives the model something to retrieve when asked in it. Genuine third-party presence in that language - real participation in community forums and question threads, not seeded reviews - adds language-matched mentions the model can find. A clearly connected professional identity across profiles, so the model can confirm it is the same attorney whether it encounters the name in one language or another, reduces the chance the firm is dropped for lack of corroboration. See also choosing an immigration attorney as a foreign founder for how this looks from the client side.

None of these guarantees appearance in any given answer. AI engines do not publish their selection logic, and no honest provider can promise a specific result. What these factors do is change the evidence the model has to work with in the language where the firm is currently thin.

How to check this yourself in five minutes

Pick the question a real client would ask: an attorney for your visa specialty, who speaks your community's language. Ask an AI engine that question in English, in a normal consumer session, and note every firm it names. Then ask the equivalent question in the community language, in a fresh session, and note every firm it names there. Compare the two lists.

If your firm is in both, the language promise has a path to the people it is meant for. If it is in one and not the other, you have found the specific, checkable gap that a single-language check would have hidden - and you now know which language to work on first.

FAQ

Does the language of the question really change which firms an AI recommends?

Yes. The answer to a question asked in English is assembled from different training data and different retrieved pages than the same question asked in another language. Because the underlying material differs, the list of firms can differ too. A firm can appear in one language version and not the other.

My firm appears in ChatGPT when I check in English. Is that enough?

It confirms the model can name you in the language most of the web is written in. It does not confirm you appear when a client asks in your community's language, which is a separate answer built from separate material. If your positioning depends on serving that community, the English check alone leaves the more important question unanswered.

Why would a large English-only firm appear in a non-English answer instead of a smaller firm that speaks the language?

AI answers are assembled from whatever the model has seen and can retrieve, not from a verified directory of language skills. A large firm with a heavy web presence can be surfaced for a language query simply because it has more citable material overall, even without a language specialty, while a smaller specialist with a thin footprint in that language has little for the model to find.

Can I guarantee my firm will appear in the AI answer if I fix this?

No, and any provider who promises a specific AI result is overpromising. AI engines do not disclose their selection logic and their answers shift over time. What is possible is to give the model more language-matched, verifiable material to work with in the language where the firm is currently thin, which changes the inputs rather than guaranteeing the output.

Which language should I focus on first?

The one your positioning depends on and where the gap is widest. If you built the practice on serving a specific community, that community's language is where an absence costs you the most, because it is exactly the client for whom the language was the deciding factor.