Yes. FTA Global’s Context Report, published in September 2026, ran identical buying questions across 10 cities and seven languages. Nearly three of the top five brands changed once the buyer moved away from New York and English.
GEO, or generative engine optimization, is the work of getting AI tools like ChatGPT and Gemini to name your brand in their answers. (Our GEO explainer covers the basics.) You have to optimize market by market.
Buyers act on these answers. A March 2026 report by EMARKETER and Publicis Commerce asked US AI users about switching. Nearly half, 49%, said they were likely or very likely to try a different brand if an assistant suggested one. The survey covered 1,179 US adults who had used AI to research a product in the past month.
If the assistant names your rival in Mumbai and you test from New York alone, you won’t see that loss.
Key Takeaways
- Nearly 3 of the top 5 AI-recommended brands change when buyers leave New York and English (FTA Global, 2026).
- In a 2026 accounting-software test, language decided whether local brands appeared. Location decided which country’s brands got named.
- In our July 2026 run, ChatGPT cited a Google Business Profile for every Bangalore agency it named. In four other cities, it cited none.
- Two runs of one prompt matched less than 1 time in 100 (SparkToro, 2026). Test each market in its own language, many times.
How different are AI brand recommendations from one country to the next?
Most brands AI names are local to one country. In FTA Global’s September 2026 report, 159 of the 206 brands it recorded appeared in a single country. That’s 77% of the pool.
The study asked buying questions in 10 categories, including CRM, payroll, payments, web hosting and HR software. It covered 10 cities and seven languages. Every question went to ChatGPT and Gemini five times, for 1,200 answers in all.
Brands shifted most in money categories. For payroll, accounting, payments and HR, an average of 4.4 of the top five brands changed across markets. For everyday tools like video conferencing and cloud storage, the average was 1.5.
Four examples from the report:
- Payments, New York vs Mumbai. New York returned Stripe, Square, PayPal, Helcim and Shopify. Mumbai returned Razorpay, Cashfree, PayU, PhonePe and Instamojo. Not one brand appeared in both lists.
- HR software. Not one of the top US brands appeared in any other country tested. All five of the top brands changed against the New York answer.
- CRM, Paris. On the same question, Gemini placed Sellsy first and ChatGPT placed Pipedrive first. For the same question in the same city, the two engines shared about 3.2 of the top five brands.
- Mumbai, three languages. The same Mumbai buying question in English, Hindi and Hinglish returned Indian brands every time. Paytm replaced one brand in the Hindi results.
FTA Global launched its own visibility metric, Context Authority, alongside this report, so it has a stake in the findings. FTA says it left its clients out and took the questions from neutral sources rather than brand websites.
FTA says its full report, free with an email sign-up, includes the complete method and data. Check it before you quote exact figures.
What did ChatGPT name when we searched one term in 10 cities?
The brands changed, and so did the evidence ChatGPT leaned on. In our July 2026 test, ChatGPT cited a Google Business Profile for every Bangalore agency it named. In San Francisco, Singapore, Sydney and Nairobi, it cited one for none. The query was “GEO agency [city]”, asked in English across 19 cities.
We’ve pulled out 10 of those cities here, two per region. Each answer came from one query, in one session. “Positioning guessed” means ChatGPT described an agency’s GEO focus as its own read, with no source to point to.
| Region | City | Agencies named | Examples ChatGPT named | Business Profile cited | Dedicated GEO page cited | Positioning guessed |
|---|---|---|---|---|---|---|
| India | Bangalore | 4 | OneCity Technologies, Crimson Salt, Brandstory Solutions, Digimark Agency | 100% | 1 of 4 | 50% |
| India | Mumbai | 6 | Jump Content, Bricks Media, Rajkar AI SEO Agency, Eiosys | 50% | 1 of 6 | 33% |
| North America | New York | 6 | Geolign, The GEO Agency, Infidigit, iPullRank | 17% | 2 of 6 | 17% |
| North America | San Francisco | 5 | Geolign, SearchTides, PipeRocket Digital, Single Grain | 0% | 1 of 5 | 80% |
| Europe | London | 7 | GA Agency, The GEO Agency, GEOQ, Propeller | 57% | 4 of 7 | 14% |
| Europe | Paris | 5 | Optimize 360, NEXUS GEO, Matterz, AdsMirable | 40% | 3 of 5 | 0% |
| Asia-Pacific | Singapore | 5 | AI Studio, Rankpage, First Page Digital, OOm | 0% | 2 of 5 | 60% |
| Asia-Pacific | Sydney | 5 | StudioHawk, Online Marketing Gurus, Supple Digital, Safari Digital | 0% | 0 of 5 | 100% |
| Middle East and Africa | Dubai | 5 | SEO Tech Experts, Royex Technologies, GCC Technologies, BeOnTop | 20% | 3 of 5 | 40% |
| Middle East and Africa | Nairobi | 4 | Dotsavvy, MediaForce Communications, Urbanacle, Bluegift Digital | 0% | 0 of 4 | 100% |
You’d expect a city-named query to return local firms, and it did. Only Geolign (New York and San Francisco) and The GEO Agency (New York and London) crossed cities. The bigger difference was in what ChatGPT said it had read.
Three patterns:
- The Business Profile is an India pattern. In Bangalore, ChatGPT cited one for every agency. Four of the ten cities cited none. Across all four Indian metros in the full run, the rate was 86%, against 18% everywhere else.
- Dedicated pages went with less guessing. Paris and London, where most agencies had a dedicated GEO page, had the lowest guess rates. Sydney and Nairobi had no dedicated pages cited, and ChatGPT guessed every agency’s positioning.
- ChatGPT drew on different assets in different markets. In India, agencies wrote “GEO” straight into their Business Profile, and ChatGPT repeated it. In Paris and London, most named agencies had a dedicated GEO page for the model to read. London agencies often had a Business Profile too.
Our finding: Across 94 agency listings in July 2026, ChatGPT guessed an agency’s positioning 66% of the time when it had no dedicated GEO page. When a page existed, it guessed 0% of the time. Some of that zero is built in, since a page gives the model a source to state. In this run, a listing got agencies named, and a dedicated page gave ChatGPT their own description to repeat.
Three caveats. ChatGPT reported its own sources after the fact, so the source columns aren’t server logs. Each city’s percentages rest on four to seven listings from a single run, so treat them as signs of direction. And our agency appears in the Bangalore results, named from a Business Profile alone.
Read: A Google Business Profile Moves ChatGPT in India and Almost Nowhere Else
Does the language of the question or the location of the buyer drive the difference?
Both, but they do different jobs. In August 2026, Dmitrij Żatuchin isolated the two in an arXiv preprint, The Language of the Question Selects the Market. Query language decided whether local suppliers appeared. When the question matched the country’s official language, a global supplier appeared in 1 of 24 runs.
The probe ran 234 runs against logged-out ChatGPT and the OpenAI API on 29 and 30 August 2026. It used connections from four countries, six query languages and six identical runs per cell. The main test category was accounting software for freelancers, with project management software as a control.
- Oslo, asked in Norwegian: Fiken, Tripletex and Conta. No global product.
- Oslo, asked in English: FreshBooks, QuickBooks, Wave and Zoho. Fiken survived in one run of six.
- Istanbul and Tallinn, asked in English: no local supplier in any of the 12 runs.
- Berlin, asked in Turkish: German suppliers, written in Turkish.
- Tallinn, asked in Russian: an Estonian supplier in 4 of 6 runs, and a global one in all six.
The engine read the query language to decide whether to localize, and the connection’s location to decide which country’s market to serve. A Turkish speaker in Berlin got German suppliers, in Turkish.
If you check in English alone, you may be measuring a market you don’t compete in, and you won’t see your local rivals.
The study also ran a control. In project management software, no domestic supplier appeared in any language, even in Germany, where such suppliers exist. Żatuchin puts this down to regulation: accounting, with its VAT and e-invoicing rules, localizes, and project management doesn’t.
Our read: Two independent 2026 studies point the same way. In FTA Global’s data, money categories moved 4.4 of five brands and everyday tools moved 1.5. In Żatuchin’s preprint, accounting software localized and project management didn’t. If your category is tied to local tax or finance rules, expect local answers.
Treat the preprint with care. It isn’t peer-reviewed, it used one model and one prompt per category, and its author runs a company that sells AI-visibility measurement.
Why does the same question return different brands inside one country?
Because every AI answer is a fresh draw. FTA Global’s September 2026 study held the city and language fixed and asked each question five times. The same top brand held in 35% of cases. On average, two different brands took the number one spot across five runs.
SparkToro’s larger study, published in January 2026, found the same pattern at scale. Its 600 volunteers ran 12 prompts a combined 2,961 times. ChatGPT and Google’s AI had less than a 1-in-100 chance of returning the same brand list twice. The same list in the same order came up less than 1 in 1,000 times.
Two limits apply. The volunteers used their own settings and locations, so the figures mix run-to-run noise with differences between users. And SparkToro ran the study with Gumshoe.ai, which sells AI tracking tools. Rand Fishkin disclosed the partnership.
Some brands hold steady through the noise. When 142 people wrote their own headphone prompts, Bose, Sony, Sennheiser and Apple appeared in 55 to 77% of 994 responses. Each topic has a core set of brands, and their order changes from run to run.
Treat any single screenshot as one draw. Track how often your brand appears across many runs, in each market.
One more variable is the user. In September 2026, Cassie Wilson Clark and Joao da Silva published The Personalization Gap, a UK study of 8,609 responses from ChatGPT, Gemini, Claude and Perplexity. What an AI knew about a user changed which brands it kept recommending.
Geography was the clearest effect. In accounts primed with a user history, UK and EU brands’ share rose by 6.6 percentage points. Gemini and ChatGPT shifted most. Claude’s recurring brands didn’t move beyond normal noise.
The authors say this doesn’t prove that localizing content improves visibility. Both work in AI-visibility consulting and measurement.
Why do the sources AI reads change by country?
Because the brands an AI names follow the pages it reads, and those pages are local. In FTA Global’s September 2026 report, local pages supplied 42% of cited sources in Germany and 12% in the UAE. Those sources included local review sites, media and company pages.
If local pages supply 42% of German citations, a site with no German presence competes for the other 58%. To compete for those German citations, you need local coverage.
Buyers in different regions also use different engines.
- Google. AI Overviews reached India and the UK on 15 August 2024, 93 days after the US launch. Nine continental European countries, from Germany to Switzerland, waited 316 days. France got them on 22 July 2026, more than two years after the US. Google shows these features by the searcher’s physical location, so a test from the wrong country can show you nothing.
- China. ChatGPT doesn’t operate in mainland China, so its results tell you little there. In CNNIC’s June 2025 survey, 72.2% of generative AI users had used ByteDance’s Doubao in the past six months. DeepSeek came second at 62.0%. The order has shifted since. QuestMobile’s June 2026 app data has Doubao first with 382 million monthly users. Alibaba’s Qwen is now second with 167 million. DeepSeek is third with about 130 million, down 20% in a year.
- South Korea. Naver held a 63.8% search share in March 2026, by InternetTrend’s count, and its AI leans on Naver’s own sites. StatCounter puts Google ahead across all devices, at 46.6% to 43.7% in August 2026. But StatCounter counts clicks from search engines to outside sites, so it can’t see searches answered inside Naver. Naver disputes its figures. Name your source when you quote Korean share.
Read: International GEO Strategy: Does GEO Work the Same in the US, Europe and Asia?
How should you test and track AI visibility country by country?
Test each market on its own terms, and run every prompt many times. In SparkToro’s data, two runs matched less than 1 time in 100. One run can’t tell you where you stand. A fair test needs local language, a local connection and enough repeats to read a rate.
- Pick your markets and categories. Start with the countries that bring revenue. Note whether your category is tied to local rules like tax, payroll or finance. Those categories localized most in the studies above.
- Write the questions in your buyer’s language. Translate, then ask a local speaker to rewrite them the way buyers ask. Test mixed-language versions too, like Hinglish.
- Run them from inside the country. Use a local connection or a regional tracking tool. An API call carries no user geography.
- Repeat each prompt many times. SparkToro’s Rand Fishkin suggests at least 60 to 100 runs per prompt. No study has yet settled the right number, so adjust once you see how much your category moves.
- Record your visibility rate. Count the share of runs that name you, and list which competitors appear beside you.
- Read the sources behind the answer. Note which local sites, directories and Business Profiles the engine cites, then compare them with what your own pages say.
- Test more than one engine. In China and Korea, test the local engines first.
- Fix the gaps with local assets. Publish one clear page per market that says what you do in plain words. Add the local listings and local-press coverage the engines already read.
Log every run so you can repeat it next quarter. A simple template:
| Date run | Region | City | Language | Prompt | Runs | Brands named | Share of runs naming you |
|---|---|---|---|---|---|---|---|
| 2026-11-10 | India | Bangalore | English | “GEO agency Bangalore” | 60 | 8 | 60 (position changes) |
| 2026-11-10 | Europe | Paris | French | Agences GEO à Paris | 60 | 8 | Zero |
For the audit step, see AI Visibility Audit: What It Is and How to Run One. For page-level fixes, read How to Do a Content Gap Analysis and How Do You Get Your Brand Featured in Comparison Listicles?
Where should you start?
Each market gets its own AI brand list. In the clearest controlled test we found, language decided whether local brands appeared and location decided which market got named. The engine you ask changes the list again. A US-only or English-only check can look healthy while you lose buyers elsewhere.
Pick your next market. Write the buying questions in its language, run them from inside the country, repeat them, and read the sources behind the answers. Then build the local pages and listings those sources reward.
Crimson Salt’s AI Visibility Audit checks how AI tools describe your brand, one country at a time. If a new market is in your 2027 plans, write to [email protected].
Sources
- Social Samosa, Location and language reshape brand recommendations on AI platforms: Report, published 17 September 2026. FTA Global’s Context Report: city study, payments, HR, CRM, engine overlap, repeat-run, language and local-source figures (Charts 1 and 5). Cross-checked against FTA’s own report page, The Context Report, which gives 1,200 tests, the 77% single-country share and the 35% repeat figure. FTA’s materials disagree on the city count: the report page body says 10 cities, while its page summary and press release say 12. Our reading, not FTA’s: 1,200 answers equals 12 city-and-language combinations × 10 categories × 2 engines × 5 runs, which fits 10 cities if Mumbai counts three times (English, Hindi, Hinglish).
- Dmitrij Żatuchin, arXiv, The Language of the Question Selects the Market (arXiv:2608.30052), submitted 30 August 2026, data collected 29 to 30 August 2026 (Chart 4). Figures checked against the full PDF.
- Rand Fishkin, SparkToro, NEW Research: AIs are highly inconsistent when recommending brands or products, published 27 January 2026. Data collected November and December 2025. Also reported by Search Engine Journal, AI Recommendations Change With Nearly Every Query, 30 January 2026.
- EMARKETER, When an AI chatbot recommends a different brand, nearly half of active users say they’ll try it, citing EMARKETER and Publicis Commerce, LLM Usage and Trust in the Shopping Journey, March 2026. Base: 1,179 US adults 18+ who had used an AI assistant or chatbot to research a product in the past month, surveyed in January 2026.
- EIN Presswire, New Study Finds AI Personalization Can Change Which Brands ChatGPT and Gemini Recommend, published 23 September 2026. Press release for The Personalization Gap by Cassie Wilson Clark and Joao da Silva. Figures checked against Clark’s study write-up, 21 September 2026; full study DOI 10.5281/zenodo.22161044. The study held location fixed (UK IP addresses) and varied account history.
- Crimson Salt, A Google Business Profile Moves ChatGPT in India and Almost Nowhere Else, published 21 July 2026, updated 5 October 2026. Source of all city-level figures in the 10-city table and Charts 2 and 3. ChatGPT’s source column is self-reported.
- Crimson Salt, International GEO Strategy: Does GEO Work the Same in the US, Europe and Asia?, published 5 October 2026. Naver’s 63.8% checked against The Korea Times, 16 April 2026. Rollout dates checked against launch coverage: India and UK, 15 August 2024; nine European countries, 26 March 2025 (PPC Land); France, 22 July 2026 (PPC Land). Day counts are ours, from the 14 May 2024 US launch.
- China Internet Network Information Center (CNNIC), 生成式人工智能应用发展报告（2025） (Generative AI Application Development Report 2025), October 2025, Figure 2. Survey data to 30 June 2025, base 515 million generative AI users. CNNIC defines the usage rate as the share of generative AI users who used a product in the past six months. Other products: Tencent Yuanbao 16.5%, Kimi 16.4%, ERNIE Bot 12.2%. These figures are often credited to CNNIC’s 57th Statistical Report (February 2026, 602 million users), but that report has no product-level table.
- QuestMobile, 2026 AI application market half-year report, July 2026, as reported by 36Kr. Monthly active users of AI-native apps, June 2026. Monthly users and CNNIC’s six-month usage rate measure different things, so the two sets of figures aren’t directly comparable.
- FTA Global, press release for The Context Report, 17 September 2026, as syndicated by Newspatrolling. FTA’s own wording for the 4.4 and 1.5 figures (“across markets”), the HR finding and the 42% German-source figure.
- Google AI feature availability by physical location: Google Search Help forum, 22 May 2026, as reported by PPC Land. The statement came from a volunteer Product Expert, not Google staff.
- Naver AI Briefing sourcing: Naver’s launch announcement (24 March 2025) lists blogs, cafes, places and shopping among its sources. A 2026 study by Jeongmaek Marketing Lab, reported by Digital Today, found Naver Blog supplied 46.5% of AI Briefing citations.
- StatCounter, Search Engine Market Share in Republic of Korea, August 2026, all platforms.
