Does ChatGPT Recommend Different Companies in Different Countries? What We Found, and What to Check for Each Market

1. The short answer: both matter, and in our test language mattered most

We asked the same business questions in three ways and counted how many of the recommended companies stayed the same. Read each row as: out of every 10 companies in the first list, how many also appeared in the second list.

So the answer is yes: ChatGPT and other AI engines recommend different companies in different countries, and in our test they changed even more when the buyer typed in their own language. Being recommended in one market tells you little about the next one.

Why it happens: OpenAI's help page on ChatGPT search says it "may use an approximate location based on your IP address to provide relevant local results", and that ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to search providers. Two people typing the same question in two countries can therefore trigger different searches and read different sources. A question typed in Bahasa Indonesia leads the engine further towards Indonesian sources and Indonesian companies.

2. How we tested it

We wrote ten questions a business buyer would ask, two in each of five categories, without naming any country: commercial interior design, PR agencies, corporate gifting, tech recruitment and CRM implementation partners. For example: "Which commercial interior design firms would you recommend for a new office fit out?"

We sent each question through DataForSEO, with the asking country set on every request. For ChatGPT we used OpenAI's gpt-4o-mini model with web search switched on, which is the API version of ChatGPT search; the ChatGPT app itself may use a newer model. For Perplexity we used its Sonar model, and the third engine was Google AI Mode. Each question ran three times per engine and country. That gave 360 answers, of which 355 returned successfully.

A language model then listed the companies each answer recommended, we merged spelling variants of the same name and checked a sample against the answers, and for each question and engine we compared every company named across the three runs in one market with those named across the three runs in the other. No company from the test is named here. It is a small test of ten questions (two per category) run on one day, so treat the category figures as a direction to test against your own market.

3. Singapore and Hong Kong got different shortlists

AI answers change a little even when nothing else changes: when we asked the same question again in the same country, about 73% of the companies were named again, roughly 7 out of 10. That is the normal level of change. Moving the same English questions from Singapore to Hong Kong pulled that number down to 42%, roughly 4 out of 10. Each row below shows, out of the companies that engine recommended in Singapore, the share it also recommended in Hong Kong:

How to read the ChatGPT row: if ChatGPT recommended 10 firms to a buyer asking from Singapore, about 3 of those firms also appeared when we asked the same question from Hong Kong (across three tries), and the other 7 were replaced by different firms. ChatGPT and Google AI Mode localised the most. You can see it in the answers themselves: none of our questions named a country, yet 67 of 89 answers asked from Singapore used the word "Singapore", and 55 of 90 answers asked from Hong Kong used the words "Hong Kong".

4. Local services diverge completely

How much the list changed depended on what was being bought. Each row shows, out of the companies recommended in Singapore for that category, the share also recommended in Hong Kong:

For commercial interior design, a service delivered on site, the two markets shared no firms at all. Categories where large regional or global firms operate kept about half their names. If your business is delivered locally, treat each market as a separate contest; if you sell regionally, expect a large part of the shortlist to change.

5. In our test, language changed the list even more than location

Here the country stayed the same and only the language changed. Asked from Indonesia, the English and Bahasa Indonesia versions of each question shared only 15% of their recommended companies (52 of 339): out of every 10 companies in the English answers, only 1 or 2 also appeared in the Bahasa answers. Compare that with about 7 out of 10 when nothing changes. The engines agreed: ChatGPT 15%, Perplexity 16% and Google AI Mode 15%.

A real example from the test

Here is one question from the test, asked to ChatGPT from Indonesia on the same day, once in English and once in Bahasa Indonesia. We have hidden the firm names.

Asked in English: "Which PR agencies are best for a B2B technology company?"

"Here are several PR agencies renowned for their expertise in the B2B tech sector:" Firm A (Amsterdam, Netherlands), Firm B (London, UK), Firm C (Boston, USA), and five more firms in the United States.

Asked in Bahasa Indonesia: "Agensi PR mana yang terbaik untuk perusahaan teknologi B2B?" (the same question)

"Memilih agensi hubungan masyarakat (PR) yang tepat untuk perusahaan teknologi B2B Anda di Indonesia sangat penting..." (in English: choosing the right PR agency for your B2B technology company in Indonesia is very important). Firm D (Jakarta, Indonesia), Firm E (Jakarta, Indonesia), Firm F (Jakarta, Indonesia), and three more firms in Jakarta.

Same question, same country, same day. The English answer never talked about Indonesia and recommended eight firms in Europe and the United States. The Bahasa answer talked about Indonesia in its very first sentence and recommended six firms, all in Jakarta. Not one firm appeared in both answers.

The rest of the test leaned the same way: about half of all English and Bahasa answer pairs (42 of 85) shared no firm at all. The table counts it for all 88 answers in each language: "Answer talked about Indonesia" means the answer used the word Indonesia or Jakarta, even though the question never named a country; "Local websites used" is the share of linked websites with an Indonesian .id address.

The Bahasa questions pulled the answers towards local companies and local sources. About half the English answers from the same country did not mention Indonesia at all. A brand that appears in English answers but has no presence in its buyers' language can be invisible to the buyers who ask in that language.

Other research shows the country also changes how often brands are named at all. In a Semrush study across 14 countries, brands were named in 50% of AI answers in India and Sweden but in only 18% to 22% in Italy, Brazil and the Netherlands.

6. What to check for each market

Treat each market and language as its own prompt set and its own result. Five steps:

1. List the markets that matter. The countries where you sell and want new buyers, and the languages those buyers search in. A market you do not serve does not need tracking.
2. Build a prompt set per market, in the buyer's language. Use that market's Search Console queries, sales conversations and local forums, as in our guide to finding the AI prompts your buyers use. Translating an English list afterwards produces questions nobody asks.
3. Ask from inside the market. Because ChatGPT can use your approximate location, checking from your own office shows your own market's answer. Use a tracker that sets the country, or ask a colleague in that market to run the questions.
4. Run each question more than once. In our test, about a quarter of the companies named in one run were missing from the next run in the same market, so one answer is a sample and three runs are a picture.
5. Record who is named instead of you, per market. The competitors and the cited sources differ by market, and so does the fix.

7. How to show up in each market

When a market's answers leave you out, look at what they cite. In our test the local language answers leaned on local sources, so the work is usually local too:

• A page for each market you serve. A service page scoped to that market, with local clients, local pricing or terms, and the market named plainly, gives the engine a page that fits a buyer's question from that country.
• Content in the buyer's language. Where buyers ask in their own language, publish the key pages in it, written for that market rather than translated word for word.
• Local sources that name you. Local directories, associations, press and comparison articles in each market. A firm known in Singapore through Singapore sources is not yet known in Hong Kong.
• The same facts in every market. Office locations, services and names must match across markets and languages, or the engines will mix them up. Our guide on what ChatGPT says about your brand covers how to make those facts consistent.

The same logic covers markets you do not want: state the countries you serve on your site and profiles, so the engines describe your coverage correctly.

8. Want each market checked for you?

Our free AI visibility tracker lets you choose your market, then tests 30 to 50 of your buyers' questions in their language across ChatGPT, Gemini, Perplexity and Google AI Overview, and shows who gets named and cited. If you sell in several markets, book a 30 minute call. In our work under The Novastacks Methodology, the buyer question set is configured per market and per language, so each market gets its own questions and its own measurement.

Frequently asked questions