By Tessar Napitupulu, Founder & CEO, PT Arfadia Digital Indonesia · July 2026 · 10 min read
Ask an AI assistant which neighbourhood suits a family in South Jakarta and it will confidently name three or four areas. Ask it to name a specific project within one of those areas and it usually cannot, not because no good project exists there, but because the AI has no structured, citable content connecting that project to the neighbourhood characteristics it just described. This is the single most common frustration developers and agents raise about AI visibility, and it has a specific, addressable cause.
Why does AI answer the neighbourhood question confidently but go vague on the project question?
Because the two questions draw on entirely different kinds of source material. A "which neighbourhood" query is answered from broad, general knowledge about an area, the kind of content that exists in reasonable volume across travel guides, expat forums, and news coverage. A "which specific project" query, especially one with real constraints attached such as a budget, a unit size, or proximity to a named school, requires structured, specific, current content that AI systems can extract and match against those constraints. Most developer content does not exist in that form. It exists as marketing copy optimised for a human scrolling a brochure, not as structured data an AI system can pull a specific answer from.
Does adding more constraints to a query actually change what gets cited?
Yes, meaningfully. A generic query like "best area for families in BSD" pulls from broad neighbourhood-level content and tends to surface area names rather than specific projects. A constraint-rich query, "3-bedroom house near an international school in BSD, budget 2 billion," is a fundamentally different retrieval task, one that rewards content carrying exactly those structured attributes: named school, distance, price band, unit configuration. Projects with this structured content available have a genuine opportunity to be named in response to the second query type; projects without it are invisible to it regardless of quality, because the AI has nothing specific enough to match against the buyer's stated constraints.
Does a strategy built for ChatGPT automatically carry over to Perplexity and Gemini?
No, and this is one of the most consequential findings in current GEO research. ChatGPT dominates Indonesia's AI-driven web traffic, drawing roughly 80.6% of AI referral traffic and ranking as the country's 4th most-visited website overall, with Perplexity a distant second at around 15.03%, per DataReportal's 2026 Indonesia report. But dominance in traffic share does not mean dominance in citation overlap. Building for ChatGPT first is a defensible prioritisation given its traffic share, but treating it as the only platform that matters leaves real, if smaller, opportunity on the table with Perplexity and Gemini, and the near-zero overlap in what each engine actually cites means genuinely separate testing, not just separate awareness, is required.
How does a brand-new project get cited by AI when it has zero history?
Through entity transfer, not fabrication. A project launching this month has no reviews, no track record, and no independent coverage yet, which is a real disadvantage against an established competitor with years of accumulated signals. The fix is not to invent history that does not exist. It is to deliberately transfer credibility from sources that do exist: the developer's own track record on prior projects, neighbourhood-authority content the developer has already published and that the new project can link into, and a structured pre-launch seeding programme covering press coverage and directory listings in the 60 to 90 days before launch, so that by the day the project goes live, there is already a real, verifiable trail for an AI system to draw on.
A cold-start entity-transfer sequence for a new project
Is AI citation genuinely winnable, or does it just concentrate among a handful of already-dominant players?
It concentrates heavily, at least in the markets where this has been studied in depth, and understanding that concentration is important for setting realistic expectations rather than either giving up early or expecting rapid, easy results. Research into US real estate agent visibility found that only around 8.4% of agents appear in any AI-generated answer at all, and that the top 1% of agents who do get cited capture roughly 47% of total citation share among those who appear. That is a "winner take most" dynamic, not a level playing field, and it has not been formally studied for the Indonesian market specifically. What it does suggest directionally is that the gap between a project or agent that invests properly in structured, citable content and one that does not is likely to be large, not marginal, once AI citation becomes a more significant channel here. Being early and being structurally correct matters more in a winner-take-most dynamic than it would in a more evenly distributed one.
A useful general benchmark for judging your own citation presence, drawn from broader GEO research rather than real estate specifically, treats citation rates below roughly 8% to 15% of relevant queries as a minimal or negligible presence, a range of 20% to 30% as evidence of a strategy gaining real traction, and 40% or higher as approaching category leadership. These thresholds were not derived from Indonesian real estate data specifically and should be treated as a rough orientation rather than a precise target, but they are useful for interpreting your own weekly monitoring results: going from being named in one out of ten test queries to three or four out of ten represents genuine progress worth recognising, even well before you would expect to dominate a category outright.
Is there a technical reason a site might be invisible to AI crawlers entirely, separate from content quality?
Yes, and it is worth checking before assuming a citation gap is purely a content problem. A significant number of Indonesian websites run on closed-source or template-based content management systems that ship with an overly aggressive default security configuration, one that blocks all non-Google automated crawlers by default as a generic anti-scraping measure. That default frequently blocks GPTBot, PerplexityBot, Google-Extended, and Anthropic's crawlers alongside genuinely malicious scrapers, without the site owner realising it, since the block is invisible in normal browsing and does not affect conventional Google indexing in the way it would show up as an obvious problem. A robots.txt audit that explicitly confirms these specific crawlers are allowed is a five-minute check that resolves what would otherwise be an invisible, total blocker no amount of content investment could work around. A supplementary llms.txt file at the site root, listing key pages for AI systems, is a reasonable addition to maintain, though it is worth being accurate about its current limits: neither Google nor OpenAI has confirmed that their systems actually consume it as of this research, so it should be treated as a modest, low-cost precaution rather than a primary lever.
What is the actual technical mechanism that lets AI connect a project to a neighbourhood?
The connective layer is entity schema, specifically a sameAs array that explicitly links a project's website entity to its Google Business Profile, its verified social accounts, and any press coverage or directory listings that mention it by name. Without this explicit link, an AI system has to infer that a mention of "Grand Wisata Bekasi" in a news article and the same name on a listing page refer to the same real-world entity, and inference is less reliable than an explicit, structured statement. The neighbourhood side of the connection matters too: a project page should link to, and ideally be linked from, the developer's own neighbourhood guide for that area, using Place schema with a containedInPlace relationship establishing the geographic hierarchy explicitly rather than leaving it to be inferred from prose alone.
This is why the neighbourhood-guide content and the project-listing content need to be built as a connected system rather than two unrelated content types produced by different teams on different timelines. A beautifully built neighbourhood guide that never links to the developer's active projects, and a well-structured project page that never links back to its neighbourhood guide, both leave value on the table that a properly connected pair captures.
What does a realistic weekly monitoring routine actually look like in practice?
Keep the prompt set small and consistent rather than exhaustive: five to eight realistic buyer questions per served neighbourhood, split evenly between generic phrasing ("best area for families in X") and constraint-rich phrasing ("3-bedroom near [named school] under [budget] in X"), run against ChatGPT, Perplexity, and Gemini, in both Bahasa and English where the audience genuinely searches in both. Log three things for each run: whether your project or developer is named at all, which competitors are named instead when you are not, and roughly what kind of content the AI appears to be drawing on when it does name someone, a neighbourhood guide, a listing aggregator, a news article. That last observation is often the most actionable, since it tells you what content type to invest in next rather than just whether the current investment is paying off.
Is this actually worth the investment now, or is AI search still too small a channel to prioritise?
The trajectory argues for treating it as a near-term priority rather than a future consideration to revisit later. Research tracking AI-driven search visits globally found growth of roughly 42.8% year-on-year moving into 2026, rising from an estimated 15.6 billion visits in the first quarter of 2025 to approximately 27.4 billion by the same quarter of 2026, compared with conventional Google search visits growing only around 2.4% year-on-year over the same period. That is not a small channel growing slowly alongside a dominant incumbent, it is the fastest-growing referral category in search by a wide margin, even from a smaller absolute base. Real estate's currently low AI Overview trigger rate, discussed earlier in this piece, describes where the category sits today, not necessarily where it will sit in twelve or twenty-four months, and a channel growing at this rate is not one where "we will address it once it matters more" is a comfortable strategy.
There is also a structural shift on the horizon worth watching rather than reacting to prematurely: Google's announcement of Search and Information Agents at its I/O event in May 2026 signals a direction where AI systems increasingly complete multi-step tasks, comparing options and narrowing choices, rather than simply returning a list of links for a person to evaluate manually. The specific mechanics of how this affects real estate search are not yet established well enough to build a detailed content strategy around today, but the general direction reinforces everything covered in this piece: structured, verifiable, well-sourced content that an automated system can confidently act on is likely to matter more, not less, as these systems take on more of the comparison and narrowing work themselves.
Is Indonesian portal and developer content actually accurate when AI does cite it?
Not always, and no systematic study of Indonesian neighbourhood-content accuracy in AI answers currently exists, which is itself worth noting honestly rather than pretending the question is settled. One documented, illustrative case: an AI travel-recommendation engine described Menteng, a well-connected central Jakarta neighbourhood, as having limited public transport, despite the area being served by Jakarta's MRT, the KRL Commuterline, and Transjakarta bus rapid transit. This is a single anecdotal case, not evidence of a systematic pattern, but it illustrates the underlying opportunity clearly: AI training data can be outdated or incomplete for a specific market, and current, accurate, well-sourced content is what corrects that gap for anyone willing to publish it properly.
| Query type | What typically gets cited |
|---|---|
| "Best area for families in [region]" | Broad neighbourhood-level content, area names, not specific projects |
| "[Bedrooms] near [named school], budget [amount]" | Structured listing data matching the stated constraints |
How do you actually monitor whether any of this is working?
Run a fixed set of realistic buyer questions, covering both generic and constraint-rich phrasing, against ChatGPT, Perplexity, and Gemini on a consistent weekly schedule, in both Bahasa Indonesia and English, and log plainly whether your project or developer gets named, and who else does when it does not. This is manual and unglamorous work, and formal analytics tooling for AI citation tracking in this category is still immature industry-wide, but it is currently the most reliable way to know whether structured content changes are actually translating into citations, rather than assuming a content investment is working without any evidence either way.
Frequently Asked Questions
Should we prioritise ChatGPT given how much larger its traffic share is in Indonesia?
As a starting priority, yes, given the roughly 80.6% share of AI referral traffic it commands. But given how little overlap exists between platforms in what gets cited, treat that as sequencing, not exclusivity, and plan to extend structured testing to Perplexity and Gemini rather than stopping once ChatGPT citation improves.
Is there a shortcut to getting a brand-new project cited without waiting for the pre-launch seeding process?
Not a legitimate one. Fabricated reviews, invented history, or exaggerated claims are more likely to be caught and to damage credibility than to accelerate genuine citation, particularly as AI systems increasingly cross-reference claims against independently verifiable sources.
How long does the cold-start entity-transfer approach typically take to show results?
Early citations are realistic within 60 to 90 days of consistent, structured publishing alongside the seeding programme, with a more meaningful, durable citation presence building over 6 to 12 months as the content accumulates and gets cross-referenced across more sources.
Does correcting an inaccurate AI answer about a neighbourhood actually work, or does the old information just persist?
Current, well-sourced, properly structured content genuinely can shift what gets surfaced over time, since these systems continuously incorporate newer, more authoritative sources. It is not instantaneous, but it is not a lost cause either, and it rewards being the source that took the trouble to publish accurate information first.
How do we justify this investment to a decision-maker who wants to see leads, not citations?
Frame citation monitoring as a leading indicator, not the end goal. The realistic causal chain is structured content leads to citation, citation leads to a buyer encountering your project by name during AI-assisted research, and that buyer eventually converts through a conventional channel, a form, a call, a visit. Citation tracking tells you the first link in that chain is working well before lead volume alone would reveal it, which is valuable specifically because it gives you an early signal rather than waiting months to find out an investment did or did not pay off.