The Bappebti-to-OJK Citation Gap in Crypto AI Answers

By Tessar Napitupulu, Founder and CEO of PT Arfadia Digital Indonesia, Indonesia's GEO pioneer since 2023 and author of Cited or Silent.

Ask an AI system who regulates crypto in Indonesia, and the answer you get depends less on the question than on which engine you asked and how it retrieves information. Some will correctly say the Otoritas Jasa Keuangan, OJK. Others, trained on data that predates the transition, will still say Bappebti. Both answers sound equally confident. Only one of them has been correct since 10 January 2025.

This is not a hypothetical edge case. It is a structural gap between how different AI architectures handle a specific kind of corpus event, a regulatory transition, and it creates a category of GEO risk that most content strategies never account for: not "will we be cited," but "will the citation be accurate enough not to create a compliance problem."

The Transition That Split the Corpus in Two

Before 10 January 2025, crypto in Indonesia was regulated as a commodity by Bappebti, the Commodity Futures Trading Regulatory Agency. Peraturan Pemerintah No. 49/2024 transferred that authority to OJK, reclassifying crypto as a digital financial asset rather than a commodity, under the mandate of UU No. 4/2023 on the Development and Strengthening of the Financial Sector (UU P2SK). OJK then issued POJK No. 27/2024 to establish the operating framework for this new mandate.

The framework has kept moving since. POJK No. 23/2025, issued 4 December 2025, amended POJK 27/2024 to expand the scope of digital financial assets to include derivatives and to tighten trading requirements. POJK No. 16/2025 introduced a Fit and Proper Test for controlling shareholders, directors, and commissioners across the fintech and crypto ecosystem, effective 1 October 2025, later detailed further by SEOJK 21/2025. Most recently, POJK No. 6/2026 imposed disclosure obligations on anyone communicating about financial products to the public, including crypto, extending well beyond licensed entities to influencers and content creators.

Why Different AI Engines Get the Same Question Wrong Differently

The reason this matters for GEO specifically, and not just for keeping a website's own content current, is that AI engines do not all retrieve information the same way. Some are built primarily on a training corpus with a fixed cutoff date. Others perform live retrieval, fetching and reading current web pages at the moment a question is asked, and weighting recency accordingly.

An engine with live retrieval that is pointed at a well-maintained, dated regulatory-status page has a real chance of answering correctly, because it can see today's date and today's source. An engine relying purely on a training corpus that closed before January 2025 has no such chance unless a human intervenes with retrieval-augmented context. It will state, with the same confident tone either way, that Bappebti regulates Indonesian crypto, because that was true in its training data and nothing in its architecture forces it to flag the claim as potentially outdated.

This is a corpus problem, not a factual dispute, and it will keep recurring. Every time OJK issues a new POJK, every engine relying on a training snapshot inherits a fresh version of the same lag, until its next training run catches up, if it ever fully does for a niche regulatory detail in a single country's crypto market.

The Whitelist Accuracy Gap Is a Second, Ongoing Version of the Same Problem

The regulator transition is a one-time corpus event that will eventually age out of relevance as training data catches up. The list of which exchanges are actually licensed does not have that luxury, because it changes continuously and will keep changing for as long as OJK keeps approving, rejecting, and revoking licences.

As of May 2026, OJK had licensed 26 Pedagang Aset Keuangan Digital, PAKD, including Luno's recently approved licence. That number was smaller a year earlier and will almost certainly be different again by the time this article is read. An AI engine citing an exchange as "OJK-licensed" based on twelve-month-old corpus data could be wrong in either direction: naming an exchange that has since lost its licence, or failing to credit one that has since gained it. Neither error is cosmetic. One creates a liability for the exchange being misrepresented; the other creates a missed opportunity for the exchange that earned a legitimate licence and is not getting credited for it.

A Practical Protocol, Not a One-Time Fix

Three structures address most of this risk, and none of them are complicated to build. They are simply easy to skip, because none of them show up as urgent until an AI system has already told a user something wrong.

• Canonical regulatory-status pages, updated within 48 hours of any OJK announcement, stating the current framework in dated, extractable language: "as of [date], crypto asset oversight in Indonesia is held by OJK under PP 49/2024, most recently amended by POJK 23/2025."
• Structured OJK link anchors on every page that references licensing or regulatory status, pointing directly to the live PAKD whitelist, so a live-retrieval engine reading your page also receives a real-time verification path rather than relying on your word alone.
• A public regulatory-history page, something like /regulatory-history, documenting the Bappebti-to-OJK transition, each subsequent POJK update, and your own licence renewal timeline. This gives an engine a citeable, structured history to draw from instead of reconstructing the timeline from scattered, inconsistent press coverage.

None of these are a single project with a finish line. The regulatory calendar keeps moving, most recently with the June 2026 finfluencer disclosure rule, and a protocol that is not reviewed on a matching cadence decays back into the same accuracy gap it was built to close. This is also precisely where GEO for Crypto & Web3 and SEO for Crypto & Blockchain converge: the same dated, structured content that keeps a live-retrieval AI accurate is exactly the content Google's quality raters expect from a YMYL financial page.

What a Retraining Cycle Actually Fixes, and What It Does Not

It is tempting to treat this whole problem as temporary: wait long enough, and every model eventually retrains on a corpus that includes the transition, and the Bappebti answer disappears on its own. That is only partly true, and the part that is not true matters more for a niche regulatory detail specific to one country's crypto market.

A large training run absorbs the internet's dominant signal on a topic. For something as globally discussed as, say, a major US interest rate decision, that signal is strong and consistent enough that a retrained model reliably gets it right. "Who regulates crypto in Indonesia" is not that kind of topic. It is a low-frequency, single-country regulatory fact competing for weight against millions of higher-signal training examples. A model can retrain twice after the transition date and still occasionally reproduce an outdated answer, simply because the correct answer was never reinforced by enough independent, authoritative sources in its training data to fully displace the old one.

This is precisely why a dated, structured, frequently-linked-to regulatory page matters more for a topic like this than for a higher-profile one. It is not competing to be one signal among millions. For this specific question, it can realistically be one of only a handful of credible sources an engine, retraining or retrieving live, actually encounters.

What "Dated and Extractable" Actually Looks Like on the Page

Most regulatory content fails not because it is inaccurate at the moment of publishing, but because it is written to be true indefinitely, which is a promise no regulatory statement in this sector can actually keep. Compare two ways of writing the same underlying fact:

Common, and quietly wrong within months: "Our exchange operates under Bappebti's regulatory framework, ensuring safe and compliant trading for all Indonesian users."

Structured for both a quality rater and an AI system: "As of July 2026, [Exchange Name] operates under OJK's oversight as a licensed Pedagang Aset Keuangan Digital (PAKD), following the transfer of crypto regulatory authority from Bappebti to OJK on 10 January 2025 under PP No. 49/2024. Current licence status can be verified directly on OJK's PAKD whitelist [linked]. This page was last reviewed on [date] against OJK's most recent regulatory updates."

The second version does three things the first cannot: it states a specific date so its own claim can be evaluated for currency, it names the specific legal instrument rather than a vague "framework," and it gives both a human reader and a retrieving AI system a direct, live verification path instead of asking either to simply trust the sentence. This is the difference between content that ages gracefully and content that becomes a liability the moment the next regulation lands.

A compliance team's instinct is to treat regulatory accuracy as an internal documentation problem, solved once the right disclosures exist somewhere on the corporate website. That framing misses the audience that actually matters here: not a regulator auditing your filings, but a prospective user asking an AI system a plain question and trusting the answer at face value, with no idea that the answer might be running on eighteen-month-old training data.

The user never sees your compliance documentation. They see whatever the AI says back to them. If that answer is wrong, correcting it after the fact does not un-happen the conversation that already occurred. The only lever available before the fact is making sure the correct, current, dated answer is the one most likely to be retrieved, which is a content and structure problem as much as a legal one.

Frequently Asked Questions

If an AI cites an outdated regulator for our exchange, what is our actual exposure?

Reputational and compliance exposure, not a bug you can file a support ticket about. A user who hears an incorrect regulatory claim may act on it, and correcting it after the fact does not undo the conversation. The fix is proactive: a dated, structured regulatory-status page an engine can extract correctly in the first place.

Does this problem go away once every AI model retrains after the transition date?

Only for the one-time transition event. The licence whitelist itself keeps changing continuously as OJK approves, rejects, and revokes PAKD status, so the underlying accuracy problem is ongoing, not something a single retraining cycle resolves permanently.

Should we just state our OJK licence number once on our About page and move on?

That is a start, not a protocol. A single static mention ages the same way any other undated claim does. A canonical, dated regulatory-status page, updated within 48 hours of any relevant OJK announcement, is what keeps that information reliably current for both search engines and AI systems.

Why does licence status matter more for GEO than for regular SEO?

Both matter, but GEO raises the stakes: a Google ranking that is slightly outdated is a minor inconvenience a user can verify by clicking through. An AI-generated answer often is not clicked through or verified at all. It is simply believed.

Once every model retrains after the transition date, does this problem just resolve itself?

Only partly. Large training runs absorb high-frequency global signals reliably, but a single-country regulatory detail like this competes for weight against millions of other training examples. A model can retrain after the transition date and still occasionally reproduce an outdated answer, because the correct fact was never reinforced by enough independent sources to fully displace the old one.

Is Perplexity more reliable than ChatGPT for regulatory questions?

Engines with live retrieval have a structural advantage on fast-moving regulatory questions because they can read a current, dated source at query time. That advantage only materialises if the correct, current source actually exists online for them to retrieve. Neither engine can cite a regulatory-history page that was never published.

The full framework for auditing and managing AI citation accuracy across regulatory, financial, and other high-stakes content categories is covered in Cited or Silent, free to read with a registered email. Also available on Google Play Books and Apple Books.