Why Third-Party Sites Outrank Government Websites

By Tessar Napitupulu, Founder of PT Arfadia Digital Indonesia and Indonesia's GEO pioneer since 2023.

An official government website loses to a private blog for the same query because the private page is easier for a machine to read, not because it is more accurate. Research tracking all 41 government websites in one Indonesian province-level study found an SEO-optimised imitation page consistently outranking the real official page for identical procedural queries, and AI engines are now repeating the same pattern when they answer citizens directly instead of showing a list of links.

That finding should worry anyone running digital communications for a ministry, a BUMN, or a regional government office. It should also reassure them, because the fix has nothing to do with legal authority you already hold and everything to do with structure you can change.

What 41 Government Websites Actually Revealed

The study in question tracked technical SEO factors across every government website in its sample and compared organic performance for identical citizen-service queries. One page, built by an outside party purely to test on-page optimisation technique, consistently beat the genuine .go.id source. Not because the imitation page had better information. It had the same information, formatted so a crawler could actually parse it.

Two separate factors are doing the damage here, and government communications teams tend to only see one of them. The first is technical: missing schema, missing direct answers, missing freshness signals. The second is linguistic, and it is arguably the bigger problem for citizen-facing content specifically.

Two Citizens, Two Languages, One Missed Audience

Indonesian citizens searching for a government procedure rarely type the official terminology. "Cara buat paspor online 2026" gets searched constantly. "Prosedur permohonan paspor secara daring" almost never does, even though it is the phrase a ministry's own style guide would produce. The same split shows up in code-switched, mixed-register phrasing like "gimana cara buat NPWP online 2026", which blends colloquial and formal Indonesian in a way no official document would ever write.

A government page optimised only for its own formal language is, in effect, invisible to the exact audience it exists to serve. This is not a hypothetical gap. It is measurable in search volume, and it compounds every time a citizen gives up on the .go.id page and clicks the travel blog or tax-explainer site that happened to use their words.

The seasonal pattern makes the stakes concrete. Passport queries spike every year between May and August, tied to school holidays. SPT Tahunan (annual tax return) searches spike every February and March, ahead of the April 30 corporate deadline. PPDB school enrollment searches spike every May and June. Each of these windows is predictable a year in advance, and each one is currently an opportunity for whichever unofficial site has bothered to publish a well-structured explainer before the ministry's own content team gets to it. Seasonal content planning needs to lead these dates by eight to ten weeks at minimum, given how long government content approval chains typically run, and most agencies plan reactively instead.

There is also a quieter cost to the language gap that rarely makes it into a briefing document: every citizen who lands on an unofficial explainer instead of the official page is a citizen whose only record of "what the government said" is whatever that third party chose to publish. When that page is accurate, the harm is invisible. When it isn't, the ministry inherits the reputational fallout for a statement it never made and cannot easily correct, because it doesn't control the page carrying its own procedure.

Where AI Made the Problem Worse, Not Better

Search engines at least showed ten results and let a citizen scan past the bad one. AI answers pick one narrative and present it as settled. That collapses the second-chance mechanism traditional search still offered.

A 216-answer study of AI Mode responses to Indonesian banking questions found government and regulatory sources (OJK, Bank Indonesia, LPS) cited in only 1.9 percent of relevant answers, despite being the only entities with legal authority over the topic. That figure is specific to finance, not a general government citation rate, but it is the closest measured proxy available for Indonesia, and the direction is consistent with global data: government and official sources account for roughly 8 percent of citations across broad AI Overview studies, against 11 percent for Wikipedia and encyclopaedic sources alone.

There is a second, uglier factor. Security research has found 958 out of 1,482 suspected regional government domains compromised with injected gambling content. An AI retrieval system encountering that kind of domain contamination has no reason to treat the .go.id suffix as a trust signal, whatever structural authority the registration itself is supposed to confer.

The traffic pattern underneath all of this is shifting in a direction that makes the stakes worse, not better. Analysis of AI Overview behaviour shows organic click-through rates for informational queries shrinking from roughly 15 percent of total search volume down to around 8 percent, a drop attributed directly to users reading the AI summary and never clicking through to any source at all. Embedded citation links inside those summaries fare even worse, capturing only around 1 percent of total pageviews, meaning even the sources an AI answer does cite rarely get an actual visit. Zero-click sessions, where a user reads the summary and closes the tab without opening anything, have climbed to roughly 26 percent of total sessions. Put plainly: the traditional second-chance mechanism, where a citizen who clicked the wrong result could still notice it looked off and go back to try another one, is disappearing. Whichever single source the AI happens to synthesize into its answer functions, for a growing share of citizens, as the only source they ever see.

What makes an unofficial site win that single slot comes down to a fairly mechanical contrast. A typical unoptimised .go.id page presents its procedure inside a scanned PDF running to dozens of megabytes, or embeds the actual steps inside a table rendered as an image with no alt text, or surrounds the content with script-heavy page elements that are slow and expensive for a crawler to process. Faced with that, an AI crawler frequently just skips the page rather than pay the computational cost of extracting anything from it. The competing unofficial explainer, by contrast, opens with a one-paragraph answer box, follows with a clean question-and-answer format, and carries structured data the crawler can parse in a fraction of the time. The AI model isn't making an editorial judgment that the blog is more trustworthy. It's making a purely mechanical decision about which page cost less to extract.

Structural Element	Typical .go.id Page Today	Restructured for AI Extraction
Opening content	Regulatory preamble and legal citation before any answer	40 to 80 word direct answer, regulation cited inline
Structured data	None, on the large majority of pages	FAQPage, HowTo, GovernmentService JSON-LD
Headings	Formal policy language, mirrors internal terminology	Question-format, mirrors how citizens actually search
Freshness signal	No visible update date on most service pages	dateModified within a 90-day review cycle
Accessibility	Part of 2,088 WCAG violations found across 34 provincial sites	Remediated against WCAG 2.1, same fix often improves extractability too

The Fix Is Structural, Not Editorial

None of this requires rewriting policy or changing what a ministry is legally required to say. It requires changing how that same information is packaged. A GovernmentService schema block, for example, lets a page expose the exact fields an AI system wants: the responsible agency, the eligibility criteria, the fee, the service channel, and the date the information was last confirmed. Something like this, adapted to the specific service:

{
  "@type": "GovernmentService",
  "name": "Permohonan Paspor Baru",
  "provider": { "@type": "GovernmentOrganization", "name": "Direktorat Jenderal Imigrasi" },
  "areaServed": "Indonesia",
  "dateModified": "2025-11-01"
}

Pair that with an FAQPage block for the questions citizens actually type, and the same regulatory content that currently loses to a five-year-old blog post becomes the page an AI engine has every structural reason to cite instead.

The same discipline applies when a policy changes and an old page can't simply be deleted, because deleting it throws away whatever backlink authority it accumulated. The better sequence has three concrete steps. First, update the content directly at the existing URL where possible, or if a genuinely new URL is needed, issue a permanent 301 redirect from the old address so the accumulated SEO value transfers rather than resets to zero. Second, if the old page needs to stay live for reference, inject a visible banner at the very top stating plainly that the regulation has been superseded, naming the new rule, with a direct link to the current version. Third, update the page's dateModified schema field the same day the change goes live, not on the next scheduled content review, so a crawler checking freshness signals sees the update immediately rather than weeks later. None of these three steps requires new legal sign-off beyond what the policy change already received. All three are the difference between an AI engine citing the current rule and confidently repeating one that expired months ago.

What This Means If You Run Communications for a Ministry or BUMN

The institutions best positioned to fix this are, structurally, the ones with the least incentive to notice it. A ministry's communications unit is measured on policy accuracy and sign-off compliance, not on whether ChatGPT cites imigrasi.go.id over a travel blog. That is exactly why the gap has persisted long enough for it to show up clearly in the research: nobody's job depended on closing it, while the unofficial sites had every commercial incentive to keep optimising.

Closing it does not require new legal authority or a policy change. It requires treating the citizen-service page as a piece of retrieval infrastructure, reviewed on the same cadence as any other public-facing system, with the same person accountable for whether it's actually being found. That is a workflow change, not a content-writing exercise, which is why it tends to get assigned to whoever handles the website rather than whoever handles policy communication.

It also tends to get stuck at the procurement stage before it ever reaches a content team. Digital marketing and technical remediation work for a ministry or BUMN has to move through LKPP's e-Katalog or a formal tender, and misclassifying the scope, treating it as consultancy work rather than the "Jasa Lainnya" category it usually falls under, changes both the value threshold and the selection method before a single sentence gets rewritten. Getting that classification right the first time saves months, a topic worth its own separate look at how agencies actually get hired for this kind of work.

For the fuller model behind this, including how Entity SEO and Knowledge Graph signals interact with structured data at scale, Tessar Napitupulu's Found Before They Search walks through the three-layer SEO, GEO and AEO framework this article draws on, free to read. If your agency needs a baseline audit of exactly which citizen queries currently send people past your official page, Arfadia's Government SEO service starts there, and the companion piece on what government GEO changes once AI is answering directly covers what happens after the structural fixes are in place.

---

Frequently Asked Questions

Does restructuring a page mean rewriting the actual policy content?

No. The legal content, the requirements, the fees, stay exactly as approved. What changes is where the direct answer sits on the page, whether it's marked up as structured data, and whether the headings match how citizens phrase the question. None of that touches policy substance or requires a new sign-off on the underlying regulation.

Is this only a problem for national ministries, or does it affect regional government too?

It affects both, and pemda sites often show it more severely, since regional government pages tend to have thinner technical resourcing than national ministries. The WCAG accessibility study covering all 34 provincial websites and the domain-compromise research covering regional .go.id sites both point the same direction.

How would we even know if AI is sending citizens to the wrong source right now?

The direct way is to run the actual citizen queries through ChatGPT, Gemini, Perplexity and Google AI Overviews and log which domain gets cited. There is no shortcut that avoids doing this for real, since no aggregate study measures this specifically for Indonesian government services.

Does fixing accessibility issues help with AI citation too?

Often, yes, because the underlying causes overlap. Missing alt text, poor heading hierarchy and unstructured content hurt both a screen reader and an AI crawler. Remediating for WCAG 2.1 frequently improves extractability as a side effect, even when that was never the primary goal.

Do we need to publish in English as well as Bahasa Indonesia?

Only where the service genuinely has a foreign-facing audience, such as tourism-related permits or investment procedures. For purely domestic citizen services, the higher-value fix is almost always making the Bahasa Indonesia version match how citizens actually search, not adding a second language.

---