Why AI Does Not Recommend New Brands Yet

A brand launches in Indonesia, publishes a well written website, and someone on the team asks ChatGPT to recommend suppliers in their category. The brand does not appear. Neither does it appear in Gemini, Perplexity, or Google's AI Overviews.

The instinct at that point is to publish more. More pages, more detail, more of the brand explaining itself clearly. That instinct is half right, and the half that is wrong wastes the most time, because the mechanism deciding whether an assistant mentions you does not run primarily on what you say about yourself.

This article explains what that mechanism actually appears to be, what the research says, why Indonesia is an unusual case, and what a new entrant can reasonably expect. It also states plainly where the honest limits of the answer sit.

The starting position, stated precisely

A brand entering a new market arrives with no search history, no third party coverage, no reviews and no citation footprint. Search engines have nothing to rank. Assistants have nothing to cite.

That is not a content problem, and treating it as one is the expensive misdiagnosis. Three things are missing at once, and only one of them is fixed by writing.

What the research actually says

Two findings are worth knowing, because together they explain the shape of the problem better than any vendor claim.

Models amplify existing citation concentration. Research analysing large volumes of AI generated references has found that language models systematically reinforce a Matthew effect in citation, favouring sources that are already frequently cited over newer ones. In plain terms, being cited makes you more likely to be cited.

The uncomfortable implication for a new entrant is direct. Brand owned content does not appear in third party citation graphs. Earned coverage does. Publishing more pages on your own domain improves what an assistant can read about you, and does very little to change whether the assistant considers you worth mentioning in the first place.

Content features affect citation likelihood. Separate research on generative engine optimisation found that content carrying concrete statistics, named sources and inline citations is materially more likely to be used in AI generated answers than content that asserts without attribution.

That points to a specific kind of content rather than simply more of it: verifiable, sourced, and structured so a model can extract a statement it can stand behind. Which is a different brief from most brand content, and a harder one.

Why Indonesia is an unusual case

Two facts about the Indonesian market sit oddly together, and the gap between them is the whole opportunity.

Adoption of any generative engine optimisation stands at 7 percent among Indonesian small and medium enterprises and 31 percent among enterprises, from the same survey. Both figures are low against a market where daily AI use is close to universal.

The language split that most plans miss

There is a structural feature of Indonesian search that a single language strategy handles badly.

Search intent divides between Indonesian and English, and the division is not even across categories. Technical and business to business queries skew more toward English. Consumer and local service queries skew heavily toward Indonesian. A brand publishing only in English is invisible to a large share of the intent that matters, and a brand publishing only in Indonesian misses the segment that researches in English.

The answer is both, and not as translations of each other. A translated page answers the question the original was written for, in a language where people ask a different question. The queries differ, so the content has to.

Where the sources an assistant trusts actually live

This is the part that changes what a visibility plan looks like in Indonesia specifically.

Indonesian buyers evaluate brands across search, social platforms, marketplace reviews and community discussion rather than on a company website. Search accounts for 38.3 percent of brand discovery and social ads 37.3 percent, functionally tied, while social comments and posts account for 32.6 percent. Six in ten Indonesians use social media as their primary channel for researching brands.

Which means the sources an assistant draws on when forming a view of your brand are largely not sources you control. Your website is one input among many, and in a market where community discussion carries this much weight, it may not be the dominant one.

There is a further wrinkle specific to Indonesia. A meaningful share of that community discussion happens inside WhatsApp groups, which are closed. Standard listening tools see very little of it, and neither do models. It shapes purchasing decisions while remaining largely invisible to measurement, which is worth acknowledging rather than pretending otherwise.

Two things that look like progress and are not

Because the mechanism is poorly understood, a market has grown around it, and two offers recur that are worth recognising.

Structured data as a lever. Schema markup is genuinely useful. It helps a machine understand what a page is about, and it is table stakes for a serious site. What it does not do is cause citation. Google's own published guidance is clear that structured data describes content rather than promoting it, and that it must describe content actually visible on the page. Marking up content a user cannot see breaches the guidance and risks a penalty. A proposal that presents schema as the mechanism for AI visibility has misunderstood what it does.

Directive files as a requirement. Files intended to instruct AI crawlers have become a common deliverable. They are inexpensive and harmless to add, and adding one is reasonable. Presenting one as a requirement, or as the reason a brand will start appearing in answers, is a different claim, and it is not supported. No file instructs a model to cite you.

The pattern in both cases is the same: a real technical practice repositioned as the causal mechanism. The actual mechanism, as far as the research indicates, runs through what independent sources say about you, which is slower to build and harder to sell.

How to tell whether a proposal is serious

Four questions separate a measurement approach from an assertion, and they are worth asking before signing anything.

How is a single citation counted, precisely. Does a brand mention count, or only a linked reference. Does an answer that names you without linking count. Without a stated definition, any number reported later is uninterpretable.

Which prompts are being monitored, and are they fixed. A panel that changes between reporting periods produces movement that reflects the panel rather than the brand.

What was the baseline before work started. Improvement claimed without a documented starting point is not a measurement.

What is the comparison set. Share of answers where you appear against named competitors is a meaningful figure. A count of appearances with no denominator is not.

Any provider working seriously in this area can answer all four in a sentence each. The inability to answer them is itself the answer.

What actually moves it

Four things, in rough order of how much they matter for a brand starting from zero.

Earned third party coverage. This is the one that changes the citation graph, and it is the hardest. A single placement in a publication that a model already treats as credible is worth more than a large volume of self published content, because it exists outside your domain.

Content that can be extracted. Specific, sourced, structured. If a model cannot lift a defensible statement out of a page, the page is not doing visibility work regardless of how well written it is. This is where the research on statistics and inline citations applies directly.

Presence in both languages, written natively. Not translated. The queries differ between Indonesian and English, so the content should be written for each rather than converted.

Verifiable entity signals. Consistent business information, structured data that describes what is actually visible on the page, and presence in the business databases models draw on. Unglamorous, and it establishes that the brand exists as a real entity rather than a website.

Why the timing argument is the strongest one

Every other market entry track has a regulator forcing a schedule. Product registration has a deadline. Trademark filing has a window. Licensing has a process. Visibility has none of those, which is precisely why it slips.

Three properties make that slippage expensive. It compounds, so six months started now is worth more than twelve months started later, because the early months are what the later ones build on. It cannot be bought quickly, since paid media buys attention while the budget runs whereas citation footprint persists but only accrues with time. And it runs in parallel, meaning it does not have to wait for any regulator.

Which produces an uncomfortable conclusion for anyone sequencing a market entry. A brand that completes every compliance track and remains invisible has bought the right to sell in a market that cannot find it.

Frequently Asked Questions

Why does our brand not appear in ChatGPT or Gemini answers?

Most likely because three things are missing at once: no accumulated ranking history on a new domain, no third party coverage for models to draw on, and nothing specific enough on your own site for a model to extract a defensible statement from. Only the third is fixed by publishing more.

Why is publishing more of our own content not enough?

Research on AI citation patterns has found that models systematically amplify a Matthew effect, favouring sources already frequently cited. Brand owned content does not appear in third party citation graphs while earned coverage does. Your own pages give an assistant something authoritative to read, but they do not by themselves make the assistant consider you worth mentioning.

What kind of content is more likely to be cited?

Research on generative engine optimisation found that content carrying concrete statistics, named sources and inline citations is materially more likely to be used in AI generated answers than content that asserts without attribution. That points to verifiable, sourced, extractable content rather than simply more content.

Should we publish in Indonesian or English?

Both, and not as translations of each other. Search intent divides between the two languages and the division varies by category, with technical and business to business queries skewing toward English and consumer queries heavily toward Indonesian. Translated pages answer the question the original was written for, in a language where people ask a different one.

How long before a new brand appears in AI answers?

Months rather than weeks, depending on category competitiveness and how much independent coverage already exists. It accelerates once third party sources begin referencing the brand. Anyone quoting a fixed number without knowing the category is guessing, and no platform guarantees outputs.

Is Indonesia different from other markets on this?

Yes, in an unusual way. Around 80 percent of Indonesian digital users interact with AI applications daily, and 72 percent of marketers use AI to produce content, but only 12 percent optimise so that AI can find them. Adoption of any generative engine optimisation is 7 percent among small and medium enterprises and 31 percent among enterprises. The competitive frontier is not yet crowded.

Can visibility work start before compliance is finished?

Yes, and that is the argument for starting early. Unlike entity, product and trademark work, visibility does not wait for a regulator. It compounds, so the months spent waiting are the expensive ones, and it cannot be bought quickly later.

---