What ChatGPT Knows About Your Industry (And What It Gets Wrong)

I asked ChatGPT to explain the Indonesian industrial pump market. It gave me a confident, well-structured, completely wrong answer.

Not wrong in a subtle, needs-a-footnote way. Wrong in a way that anyone who has sold a single pump in Indonesia would catch in five seconds.

Then I asked about leather craft in Bogor. Same confidence. Same structural wrongness. Then book conservation in Southeast Asia. Same pattern.

This is not a ChatGPT problem. It is a data problem. And if you understand it correctly, it is your opportunity.

The test: three industries I actually work in

I run three companies. PT Arsindo Cipta Karya sells and services industrial pumps. Hibrkraft makes handcrafted leather goods and does book conservation. Witanabe handles digital infrastructure. I am not an analyst reviewing these industries from a distance. I am in them. Daily.

So I tested ChatGPT across all three. Not with trick questions. Just normal questions a buyer, student, or journalist might ask.

Test 1: Indonesian pump industry. I asked: "Who are the main industrial pump distributors in Indonesia?"

ChatGPT listed Grundfos, KSB, and Sulzer. Those are manufacturers, not distributors. It confused two completely different layers of the supply chain. It also mentioned "local distributors like Torishima Indonesia" without noting that Torishima is a Japanese manufacturer with a local subsidiary, not an Indonesian distributor. The actual distribution landscape in Indonesia, dominated by regional dealers, system integrators, and project-based sales through companies like ours, was invisible.

Test 2: leather craft in Bogor. I asked: "What is the leather craft scene in Bogor, Indonesia?"

ChatGPT talked about Tanggulangin (that is Sidoarjo, East Java, not Bogor). It mentioned batik leather as a Bogor specialty. It is not. It referenced "traditional leather markets in Bogor" that do not exist. What does exist: a small cluster of independent makers doing bookbinding, journal craft, and restoration work. ChatGPT had no idea about any of it.

Test 3: book conservation in Southeast Asia. I asked: "Who are the leading book conservation practitioners in Southeast Asia?"

ChatGPT named the National Library of Singapore's conservation department (real) and the "Thai National Archives Conservation Unit" (plausible but vague). For Indonesia, it mentioned the National Library in Jakarta and then generated what appeared to be fictional practitioners. When pressed for names, it hedged. The actual community of conservation practitioners in Indonesia is tiny, undocumented online, and completely invisible to AI.

The pattern: confident, structured, wrong

Across all three tests, ChatGPT did the same thing. It answered with confidence. It organized its response with headings and bullet points. And it filled gaps in its knowledge with plausible-sounding information that was either wrong or fabricated.

This is not hallucination in the dramatic sense. This is what happens when a model trained primarily on English-language, US-centric data encounters a query about a market where authoritative structured data barely exists online.

As I covered in How AI Training Data Decides Who Gets Cited, LLMs generate answers from patterns learned during training. The training data for most models is overwhelmingly English-language and weighted toward the US and EU markets. Indonesian industry knowledge exists in the training data only to the extent that someone published it in English on a crawlable, high-authority website before the training cutoff.

For most Indonesian industries, that extent is close to zero.

What ChatGPT knows vs. what it gets wrong

I tested five Indonesian industries systematically. For each one, I asked five to eight standard questions a buyer or researcher would ask. Then I graded the responses.

The pattern is clear. The more globally traded and culturally documented an industry is, the better ChatGPT performs. Palm oil and batik have relatively rich English-language coverage because they are export commodities and cultural icons. Industrial pumps, leather craft, and book conservation are economically significant but locally documented, mostly in Indonesian, often not online at all.

Knowledge depth by region

This is not just an Indonesian problem. It is a structural bias in how LLMs are trained. The training data is overwhelmingly English-language and weighted toward markets that produce the most crawlable web content.

I estimated knowledge depth across four regions for the same five industries, based on response accuracy and specificity. The US baseline is what ChatGPT was trained on most heavily. Everything else falls off a cliff.

Look at the Indonesia column. Palm oil and batik score well because those industries have been documented in English by international organizations, NGOs, trade groups, and journalists. The data exists. It was crawled. The model learned it.

Industrial pumps in Indonesia: 25%. Leather craft: 15%. Book conservation: 10%. These are industries where the knowledge lives in the heads of practitioners, in Indonesian-language trade publications, in WhatsApp groups, and in project documentation that never touches the open web.

The US scores 92% on industrial pumps because every American pump manufacturer, distributor, and trade association has a website. They publish case studies, spec sheets, application guides, and white papers. That content gets crawled, indexed, and fed into training data. The knowledge ecosystem is mature.

Indonesia's knowledge ecosystem for most industries is not immature. It is absent.

Why this is a structural problem

Three factors compound the gap.

Language barrier. Most Indonesian industry knowledge exists in Bahasa Indonesia. LLM training data is 60-70% English. Indonesian content is present but underrepresented, and when it is included, it competes with English content for attention during training. A 2023 ACL study found that ChatGPT struggles to even distinguish between Indonesian and Malay, often responding to Malay prompts in Indonesian ^[1]. If the model cannot reliably separate two related languages, how much nuanced industry knowledge do you think it has absorbed?

Platform gap. Indonesian businesses are active on Instagram, WhatsApp, and Tokopedia. They are not active on the platforms that LLMs train on. Common Crawl does not index WhatsApp. Instagram posts are not in the training data. The platforms where Indonesian business knowledge lives are invisible to the training pipeline.

Publication gap. There is almost no tradition of English-language industry publication for Indonesian niche markets. No "Indonesian Pump Quarterly." No "Southeast Asian Conservation Review." The trade publications that do exist are in Indonesian, published as PDFs, distributed via email, and never uploaded to crawlable web pages. This is the data gap I explored in Industri yang Dikutip ChatGPT.

The result: when ChatGPT encounters a question about Indonesian industrial pumps, it has almost nothing to work with. It fills the vacuum with the closest thing it does know, which is the global (read: American and European) pump market. Then it projects that knowledge onto Indonesia and hopes for the best.

It is the informational equivalent of a tourist giving directions in a city they have never visited. Confident. Structured. Wrong.

The opportunity nobody is taking

Here is where it gets interesting.

If the knowledge gap exists because nobody has published authoritative, structured, English-language content about these industries, then the first person to do it becomes the training data.

Not metaphorically. Literally. The next time these models retrain, or the next time their RAG systems crawl for answers to these questions, the content they find will shape their responses. If your content is the most authoritative, structured, and verifiable source on a topic, you become the default citation.

This is not theory. I have watched it happen in real time with AI and industrial engineering content. Publish structured content about a topic where no structured content exists, and AI systems pick it up fast. There is no competition. You are not fighting for position 1 on a crowded SERP. You are filling a vacuum.

The economics are absurd. In a well-documented US industry, becoming the authoritative source means outcompeting thousands of existing publishers. In an undocumented Indonesian industry, it means being the first person to show up.

How to fill the gap: a practical framework

I am a practitioner. Not a guru. So here is what I actually do, and what you can do if you work in any Indonesian industry that AI currently misrepresents.

Step 1: Audit the gap. Ask ChatGPT, Gemini, and Perplexity the questions your customers ask you. Write down what they get wrong. That is your content roadmap. Every wrong answer is a publishing opportunity.

Step 2: Publish structured corrections. Not blog posts. Not opinion pieces. Structured content with clear headings, specific data, named entities, and verifiable facts. Use schema markup. Include tables and lists that AI can parse easily. The goal is not to rank on Google (though you will). The goal is to become training data.

Step 3: Write in English. This is counterintuitive for an Indonesian market. But the training data bias is toward English. If you want to influence what AI knows about your industry, you need to publish in the language the models understand best. Bilingual is ideal. Indonesian for your local market. English for the AI training pipeline.

Step 4: Use high-authority platforms. Do not just publish on your website. Put your industry knowledge on platforms that training pipelines prioritize. Zenodo for research documents. ORCID for your professional profile. Wikidata for your organization. Google Scholar for your publications. As I covered in How AI Training Data Decides Who Gets Cited, the platform is the signal.

Step 5: Be specific and verifiable. AI models learn to trust content that can be cross-referenced. If you write "the Indonesian industrial pump market is worth $2.3 billion annually," include the source. If you name distributors, include their locations and product lines. Specificity is the trust signal. Vague content gets treated as noise.

Step 6: Maintain and update. This is not a one-time activity. Models retrain. RAG systems re-crawl. If your content stays accurate and current while everything else stays stale, your authority compounds over time.

What this looks like in practice

Let me give you a concrete example from my own work.

When I started publishing structured content about industrial pump applications in Indonesia, the content landscape was effectively empty. A few manufacturer websites with global product catalogs. Some Indonesian-language forum posts. Zero structured, English-language content about the actual distribution landscape, pricing structures, or application patterns in the Indonesian market.

I published detailed content. Specific pump types for specific applications. Regional considerations (coastal vs. highland, chemical exposure profiles, tropical humidity factors). Named the standards (SNI, API). Included technical specifications that a real buyer would need.

Within months, AI search tools started citing the content when users asked about pump applications in Indonesia. Not because I gamed any algorithm. Because I was the only structured source that existed.

The same pattern applies to any Indonesian industry where the gap exists. And based on my testing, the gap exists in most of them.

The first-mover advantage is real

In traditional SEO, first-mover advantage erodes quickly. Someone with more backlinks, more domain authority, or more content can overtake you.

In AI training data, first-mover advantage is different. The content that exists in the training data when the model is trained becomes part of the model's foundational knowledge. Future content has to overcome that foundation. If the model learned that you are the authority on Indonesian industrial pumps during training, a competitor publishing similar content later has to fight against an established prior belief.

This advantage is not permanent. Models retrain. New data gets incorporated. But it is meaningful, especially in markets where no one else is publishing.

Right now, for most Indonesian industries, the first-mover advantage is available. Nobody has claimed it. The field is literally empty.

What this means for Indonesian businesses

Every Indonesian business owner who has asked ChatGPT about their industry and gotten a wrong answer has felt a moment of frustration. "It does not know anything."

That frustration is correct. And it is the wrong reaction.

The right reaction is: "It does not know anything, and I can be the one who teaches it."

Not through some complicated AI optimization strategy. Through the oldest knowledge-building mechanism that exists: publishing what you know, in a structured format, on platforms that matter.

The Indonesian practitioner who publishes authoritative structured data about their industry in English becomes the default source for AI. Not one of many sources. The source. Because there is no competition.

That is the gap. That is the opportunity. And right now, almost nobody in Indonesia is taking it.