Malaysia GEO: AI Answers Malay Prompts in Indonesian

By Tessar Napitupulu, Founder & CEO, PT Arfadia Digital Indonesia

Ask ChatGPT a question in clear, unambiguous Bahasa Melayu, and there is a real chance it answers back in Bahasa Indonesia instead. Not because it misunderstood the input. A controlled test found the model correctly tagged most Malay prompts as Malay, then answered in Indonesian anyway, because Indonesian is so much more heavily represented in the data these systems were trained on.

That single finding, from a study by Hiroki Nomoto at the Tokyo University of Foreign Studies, is the clearest evidence yet that Bahasa Melayu and Bahasa Indonesia, despite sharing a Malay root, occupy separate and unequal citation pools inside generative AI. For any business trying to get found, quoted, or recommended by an AI engine in Malaysia, this is not an academic curiosity. It is the single biggest structural risk in a bilingual GEO strategy.

What Nomoto's Study Actually Found

The methodology was simple by design. One hundred prompts, written unambiguously in standard Bahasa Melayu, were submitted to ChatGPT (GPT-3.5). Researchers then checked two things separately: did the model correctly identify the input language, and did it respond in that language.

The results split cleanly into three buckets. 31% of responses came back in Bahasa Melayu. 66% came back in Bahasa Indonesia. 3% mixed the two languages within a single response. Critically, the root cause was not language identification failure. The study found ChatGPT successfully tagged the large majority of inputs internally as "ms" (the ISO code for Malay). The model knew it was reading Malay. It answered in Indonesian anyway.

That distinction matters enormously for how a business should respond to it. A language-identification bug is fixable with better input parsing. A trained-in statistical bias toward the more heavily-represented language in a closely related pair is a structural property of the model itself, and no amount of correct input tagging on your end will fix it. The fix has to happen on the content side: publishing enough authentically Malaysian, natively-reviewed Bahasa Melayu material that the citation pool for Malay-language queries stops defaulting to Indonesian sources.

The Vocabulary Problem Sits on Top of the Statistical One

Even setting the training-data imbalance aside, Bahasa Melayu and Bahasa Indonesia diverge in ways that make word-for-word substitution unreliable. The two languages share a Classical Malay root but developed under different colonial administrations, Malaysia under British rule, Indonesia under Dutch, which embedded different loanword sets into everyday and professional vocabulary.

The clearest risk sits in a small set of false friends, words that look identical but mean something different, or even opposite, in each language. Percuma means "free of charge" in Malaysia, but can imply "useless" or "in vain" in Indonesian usage. Polisi means "policy" in Bahasa Melayu but "police" in Bahasa Indonesia. Kereta means "car" in Malaysia and "train" in Indonesia. Money is wang in Malaysia and uang in Indonesia; an office is pejabat in Malaysia and kantor in Indonesia. None of these are edge cases an editor would only encounter rarely. They sit in the vocabulary a marketing or compliance document uses constantly.

Tone diverges too. Indonesian business writing tends toward a more casual register even in commercial contexts. Malaysian Malay in professional and government-adjacent settings leans more formal and polite. Content written in Indonesian register and simply swapped word-for-word into Malay vocabulary reads as foreign to a Malaysian audience, in a way that undermines exactly the authority signal GEO content is supposed to build.

How ILMU 1.0 Proves the Gap Is Real, Not Theoretical

The strongest evidence that this gap has commercial consequences, not just academic interest, comes from the market's own response to it. YTL AI Labs Sdn Bhd, a subsidiary of YTL Power International Berhad, built and launched a homegrown large language model called ILMU 1.0 in August 2025, explicitly because global frontier models struggle with how Malaysians actually communicate: frequent switching between Bahasa Melayu, English and Chinese within a single sentence.

The model's own benchmark results make the underlying point directly. On the MalayMMLU test, ILMU scored 87.2%, reportedly outperforming general-purpose frontier models including GPT-5, GPT-4o and DeepSeek-V3 on Malay-specific comprehension. YTL AI Labs CEO Foong Chee Mun put it bluntly at launch: "When it comes to Bahasa Melayu, we are the best-performing LLM in the world." A vendor does not spend on building and training a dedicated national-language model unless the gap it is solving for is real and commercially significant.

This has a direct GEO implication. Where ILMU or Malaysian-language-aware systems become part of the citation ecosystem, authentically Malaysian content, reviewed by native speakers and using standard Malaysian terminology, has a growing structural advantage over content that merely resembles Malay without being built for it.

What This Means for Content Strategy in Practice

None of this means Bahasa Indonesia content is worthless for a Malaysian market entry. It can serve as a research input or a structural first draft, since the underlying grammar and much of the vocabulary genuinely overlap. What it cannot do is stand in as customer-facing, AI-citable Malaysian content without a separate, deliberate localisation pass.

A defensible content architecture treats English and Bahasa Melayu as two separate, fully-built tracks, not one track with an automated translation layer. English carries the weight for formal, procurement-grade B2B content, government-adjacent material and international audiences. Bahasa Melayu, reviewed by a native Malaysian speaker against a controlled terminology glossary, carries transactional, locally-anchored and consumer-facing queries, where the natural code-switching patterns Malaysians actually use in search (a mix sometimes called Manglish) matter more than formal correctness.

Practically, that means: maintain a standing glossary of the vocabulary pairs that most commonly get confused (percuma, kereta, wang, pejabat, and the wider set specific to your industry). Tag html lang attributes correctly (ms-MY, not id-ID) so structured data and crawlers receive an accurate signal. And most importantly, test citation behaviour directly with Bahasa Melayu prompts rather than assuming that content which performs well in English, or even in Indonesian, will transfer.

Why This Probably Matters More in Some Categories Than Others

The bias is unlikely to hit every industry with equal force, and treating it as a flat, uniform tax across all content is probably wrong. Government services, education, Islamic finance and hyper-local services plausibly have stronger, denser Bahasa Melayu source ecosystems than internationally-oriented technology or B2B categories, simply because the institutions that produce authoritative content in those categories, ministries, universities, religious authorities, publish primarily or exclusively in Malay. A query about a JAKIM halal certification process has a very different Malay-language source pool behind it than a query comparing enterprise SaaS platforms, where English-language vendor documentation and international review sites dominate regardless of the querying language.

This is a hypothesis worth stating precisely rather than asserting as settled: no controlled study in the research reviewed for this piece breaks the Nomoto-style bias-rate down by industry category. What can be said with more confidence is the mechanism that would produce this pattern: AI systems generally cite from whatever source ecosystem is deepest and most authoritative for a given query, and where that ecosystem happens to be genuinely Malay-language and institutionally anchored, the pull toward Indonesian defaults should logically weaken. Testing this by category, rather than assuming a single economy-wide bias rate, is part of what a proper Malaysian Citable Audit is for.

Malaysia Is Not the Only Country Building Around This Problem

ILMU 1.0 is not an isolated effort. At the regional level, AI Singapore has developed a family of models under the "Southeast Asian Languages in One Network" banner, SEA-LION, now at version 4.5, using a custom tokenisation approach designed specifically to reduce semantic errors across Southeast Asian language families, including both Bahasa Melayu and Bahasa Indonesia. The existence of two separate, deliberately-funded regional efforts, one Malaysian, one Singaporean, aimed at the same underlying class of problem is itself a signal worth taking seriously: the major commercial labs building general-purpose frontier models have not solved this to the satisfaction of the region's own AI researchers and institutions, which is precisely why local and regional players keep building purpose-specific alternatives.

For a business, the practical takeaway is not to wait for the general-purpose models to catch up. It is to build content and entity signals that perform well regardless of which underlying model eventually answers a given query, native Malaysian vocabulary, Malaysian institutional references, and citations from sources an AI system, whichever one it is, would reasonably weight as authoritative for a Malaysian question.

Testing the Gap Instead of Assuming It

Because this is a measurable phenomenon and not a matter of opinion, it is testable directly. A Malaysian Citable Audit, matched prompt pairs across English and Bahasa Melayu with equivalent intent, run across ChatGPT, Google AI Overviews, Perplexity and Copilot, gives a business a factual baseline instead of an assumption. A recommended first-wave design uses 150 to 300 matched prompt pairs across five to eight priority industries, with outputs coded separately for brand inclusion, source-domain inclusion, citation quality, Malaysian relevance and, critically, contamination by Indonesian vocabulary.

The controls that make an audit like this trustworthy are not exotic, but they are easy to skip under time pressure. Meaning has to be held equivalent across the language pair rather than translated literally, since a literal translation can accidentally change intent. Location needs to be specified consistently as Malaysia in both prompt versions. The same engine and account conditions should run both halves of a matched pair, within the same collection window, since AI outputs are stochastic and drift over time even for an identical prompt. Prompt type should span informational, commercial and transactional intent rather than only one, since GEO performance is not uniform across the buying journey. And because outputs vary run to run, each prompt pair needs multiple repetitions before a result is trusted, a single response is an anecdote, not a measurement.

Evaluation then codes each response across several dimensions at once: whether the brand or topic is mentioned at all, whether a citation or link is present, where in the response it appears, sentiment, and, specific to this bias question, whether the factual content is accurate for the Malaysian context or has quietly drifted toward Indonesian institutions, pricing, or terminology. This is the same audit logic already applied to the Bahasa Indonesia versus English citation gap in the Indonesian market, adapted here for a second, distinct language pair. Treating it as a hypothesis to test for every new market, rather than an assumption carried over from a different one, is the more defensible, and more accurate, way to build a bilingual GEO programme.

Frequently Asked Questions

Does this mean AI models can't understand Bahasa Melayu at all?

No. The Nomoto study found the models correctly identified Malay-language input in most cases. The issue is not comprehension, it is which language the model chooses to answer in, a choice shaped by which language dominates its training data.

Will this problem fix itself as AI models improve over time?

It may narrow as more authentic Bahasa Melayu content enters training corpora, and as purpose-built models like ILMU 1.0 gain adoption. There is no published timeline or guarantee for when, or whether, general-purpose frontier models will fully resolve it, so treating it as a present, testable risk rather than a temporary glitch is the safer planning assumption.

Can we just use Indonesian content and lightly edit it for Malaysia?

Light editing, swapping a handful of obvious words, does not resolve register differences, does not catch every false friend, and does not change the underlying training-data bias that affects which sources get cited. Genuine localisation with native Malaysian review is a materially different and more reliable process.

Does the bias run in both directions, Indonesian content getting swapped into Malay contexts and vice versa?

The documented direction in the Nomoto study is Malay prompts defaulting to Indonesian responses, consistent with Indonesian's larger footprint in training data. A symmetric, reverse-direction bias against Indonesian in an Indonesian-market context has not been documented in the sources reviewed for this piece.

How would we know if our own content is affected by this?

Run your own priority queries as actual Bahasa Melayu prompts across the major AI engines and check what language, and whose sources, come back. This is exactly what a Malaysian Citable Audit is designed to surface, rather than inferring it from English or Indonesian-language performance alone.

We go deeper on cross-market Bahasa citation testing methodology, including the full Citable Audit framework, in Cited or Silent, and apply this specific finding directly in our GEO service for Malaysia.