Global agency or local studio? For GEO, the Indonesian answer is the one to win
Ask an AI the same question in English and in Indonesian, and you can get two different answers — built from two different bodies of text. That single fact reframes a familiar choice: for an Indonesian market, the answer your buyer reads is assembled in Indonesian, from a pool a focused local studio can actually win.
AI answers are language- and market-specific. An Indonesian-language query is assembled mostly from Indonesian-language sources, which are far scarcer than English ones. A global, English-first playbook tends to optimize an answer the local buyer never reads, while a focused local studio competes in the thinner Indonesian pool where the answer is actually built. Fewer competitors, a more winnable answer.
Do AI chatbots give different answers in different languages?
Often, yes. Ask the same question in English and in Indonesian and the answers can diverge, because the model is reading two different bodies of source material.
This is not only a hunch. A 2026 study in Nature, led by researchers at UC San Diego with the University of Oregon, Purdue, NYU and Princeton, opens on exactly this point: “Ask an AI model the same political question in two different languages, and you may get two very different responses.” Across a cross-national analysis of 37 countries, the model’s answer shifted with the language of the prompt — in the China case study, the Chinese-prompted answer read as notably more favorable to China than the English one.
The mechanism the authors trace is simple, and it generalizes well beyond politics. A model learns from the text that exists in a given language. The text that exists in Indonesian is not the text that exists in English, so the Indonesian answer is a different object, drawn from a different pool.
Strip out the politics and the marketing lesson is plain. The answer an Indonesian-speaking buyer sees is assembled from Indonesian-language sources. Those sources, not the English ones, decide whether a brand is named.
Why is the Indonesian-language answer built from a thinner pool?
Because the open web a model learns from is overwhelmingly English, and Indonesian sits in the under-represented tier within it.
The raw material is lopsided. English makes up roughly 46% of Common Crawl, the dominant open dataset behind most large models, while many other languages appear orders of magnitude less often. Indonesian is one of them: as Rest of World reported, GPT, Gemini and Llama are “largely trained in English,” and Bahasa Indonesia is a low-resource language with a comparatively thin body of high-quality digital text — before you even reach the country’s 700-plus regional languages.
Read that as a problem and it is a real one. Read it as a marketer and it tends to flip. A scarce source pool is only a contested one if someone is competing for it, and in Indonesian, mostly no one is. The same answer-first page, the same clear third-party mention, carries more weight in a thin pool than in the saturated English one, because there is less to crowd it out.
Will a global agency win the Indonesian AI answer?
Usually not — because a global, English-first playbook tends to optimize for an answer the Indonesian buyer never reads.
The global model is built around the English web: English keywords, English authority sites, English-language PR, with a translation layer added at the end. That can produce excellent visibility in the English answer. But the study’s central point is that the English answer and the Indonesian answer are different objects, drawn from different sources. Translating an English page into Indonesian after the fact does not make it a native Indonesian source the way a model encounters one.
Meanwhile the things that actually shape the Indonesian answer tend to sit outside a global desk’s field of view: how a local buyer phrases the question, which Indonesian-language community threads and editorial listicles the engine leans on, what a phrase like “terbaik di Jakarta” even surfaces. Semrush’s analysis of more than 100 million citations finds that engines lean heavily on community and editorial sources; in Indonesian, those are local sources, in local language. The global agency is not bad at GEO. It tends to be winning the wrong answer.
Global agency or local studio for GEO?
For an Indonesian market, the case leans local — the local studio competes in the exact language and source pool the answer is built from.
Two playbooks, one Indonesian-language answer to win
| Global, English-first agency | Focused local studio | |
|---|---|---|
| Optimizes for | The English answer, then assumes it carries over | The Indonesian answer the buyer actually sees |
| Source pool it feeds | The crowded English-language web | The thinner Indonesian-language web, where a gap is winnable |
| Reads the prompt as | A keyword to translate | A question a local types, in local phrasing |
| Typical result | Cited in answers the buyer never reads | Named in the answer the buyer sees first |
A focused local studio starts where the answer is assembled. It writes the answer-first page in native Indonesian rather than translated Indonesian. It works to earn mentions in the Indonesian-language sources the engine cites. And it knows the real prompts — the phrasing, the “rekomendasi terbaik” constructions — because it lives in them.
Because the pool is thinner, that focused effort tends to compound faster. Princeton’s GEO research found that content backed by clear sources and statistics earned roughly 41% more citations; in a thin pool, being the one clear, well-sourced Indonesian page is a position that is easier to reach and easier to hold.
Why does this matter so much in Indonesia specifically?
Because the buyers are already on AI, already asking in Indonesian, and the click that used to be a safety net is fading.
The demand is here now. PwC finds that 69% of Indonesian workers have used AI for their role in the past year, and Indonesia’s digital-ad market sits at around US$3.41 billion in 2026. A customer is asking an engine about your category today, in Indonesian.
The old safety net — they will click through and find us anyway — is thinner than it was. A 2025 Pew study found people click a result about 8% of the time when an AI summary is shown, against 15% without one. When the answer is increasingly the destination, being named in the Indonesian answer stops being a nice-to-have.
What should you look for in a GEO partner?
Look for the partner who competes in the language and the source pool your buyer’s answer is built from — and can show you where you stand in it.
- Native, not translated. They write the answer-first page in real Indonesian, matching how locals phrase the question — not an English page run through translation.
- They know the local source pool. They can name the Indonesian-language communities, listicles and outlets the engines cite in your category, and have a plan to earn mentions there.
- They measure the Indonesian answer. They show where AI names you, and where it names a competitor, on Indonesian-language prompts — not just the English ones.
- They think in citations, not clicks. The goal is to be the named recommendation in the answer.
None of this is an argument against scale or against global craft. It is an argument about where the answer is built. For an Indonesian buyer, the answer is built in Indonesian, from a pool a focused local studio can actually compete in — the same consistency, run across every mention, that a longer narrative discipline eventually depends on. That is the case for going local.
Frequently asked questions
Do AI chatbots give different answers in different languages?
Often, yes. A 2026 Nature study across 37 countries found that asking an AI the same question in two languages can produce two genuinely different answers, because the model draws on different source material in each language. For a marketer, that means an Indonesian-speaking buyer sees an answer built from Indonesian-language sources, not the English ones.
Why are Indonesian-language AI answers easier to win?
Because the source pool is thinner. English is roughly 46% of Common Crawl, the dominant open dataset behind most models, while Indonesian is a low-resource language with far less high-quality digital text. A scarce pool is less contested, so one clear, well-sourced Indonesian page tends to carry more weight and is more likely to be the page a model names.
Will a global agency win the Indonesian AI answer?
Usually not. A global, English-first playbook optimizes the English answer and adds translation at the end, but the English answer and the Indonesian answer are different objects built from different sources. The Indonesian answer is shaped by local-language prompts, communities and outlets a global desk rarely reads.
Global agency or local studio for GEO in Indonesia?
For an Indonesian market, the case leans local. A focused local studio competes in the exact language and source pool the answer is assembled from: it writes native Indonesian, works to earn mentions in the Indonesian-language sources engines cite, and knows the real local prompts. Because the pool is thinner, that focused effort tends to compound faster.
What should I look for in a GEO partner?
One who competes where your buyer’s answer is actually built: native Indonesian content rather than translated, working knowledge of the local-language sources engines cite, measurement of where AI names you on Indonesian prompts, and a citation-first mindset rather than a click-first one.
- UC San Diego / Nature — ask an AI the same question in two languages and you may get two different answers (37-country study, 2026)
- Nature — institutional influence in LLMs: language-specific model behavior across 37 countries (2026)
- UnifiedCrawl (arXiv) — English is ~46% of Common Crawl; low-resource languages appear far less often (2024)
- Rest of World — GPT, Gemini and Llama are largely trained in English; Bahasa Indonesia is a low-resource language (2024)
- Semrush — the most-cited domains in AI answers, 100M+ citations (2025)
- Princeton — GEO: Generative Engine Optimization, ~41% lift from sources and statistics (KDD 2024)
- Pew Research — people click less when an AI summary appears (2025)
- PwC Indonesia — 69% of workers have used AI for work in the past year (2026)
- Mordor Intelligence — Indonesia digital advertising market, ~US$3.41B in 2026
Start with what you can measure
Whatever your budget, the first move is the same: see where you stand. White Wood runs a free AI-visibility report that shows exactly where AI names you — and where it names someone else — across every engine. No strings.