Asking ChatGPT About Insurance: What Actually Happens

By Tessar Napitupulu, Founder and CEO of PT Arfadia Digital Indonesia, Forbes Agency Council member, and author of Found Before They Search and Cited or Silent. Published July 2026.

Ask ChatGPT a generic insurance question and it will hedge: qualify the answer, add a disclaimer, decline to name a specific product with confidence. This is not a bug or an overcautious model. It is the same YMYL treatment Google applies to health and legal content, inherited by every major AI system through the same safety tuning, and insurance sits squarely inside that treatment because it can directly affect a person's financial security. Understanding exactly how that hedging behaves, where it is heaviest and where it is lightest, is the actual foundation of any insurance GEO strategy. Getting this wrong produces content built for a version of AI behavior that does not match how these systems actually respond.

The Hedging Is Real, Documented, and Not Insurance-Specific

Before looking at insurance specifically, it is worth establishing the pattern at the category level, because insurance inherits it rather than inventing it. Across 1,200 keywords studied by SE Ranking in September 2024, Google AI Overviews triggered on legal queries 77.67% of the time, the highest of any category tested, followed by health at 65.33%, finance at 41.67%, and politics at 16.67%, with YMYL keywords overall triggering an AI Overview 50.33% of the time. Separately, 83% of health-related keywords that triggered an AI Overview included a disclaimer along the lines of "this is for informational purposes only." Insurance was not tested as its own separate category in that specific study, but as a YMYL sibling of both health and finance, there is no reason to expect it behaves differently, and every insurance-specific source reviewed for this article treats that inference as reasonable.

These figures are global benchmarks, not Indonesian measurements, and that distinction matters enough to repeat directly: no controlled study yet exists testing how Indonesian-language insurance prompts specifically get treated across ChatGPT, Gemini, Perplexity, or Google AI Overviews. Everything in this article that draws on non-Indonesian data is flagged as such, because presenting a global or foreign benchmark as an Indonesian finding is exactly the kind of error that undermines the credibility this content category depends on.

Insurance Hedges Differently by Sub-Category, and the Logic Is Consistent

Within insurance itself, the hedging is not uniform, and the pattern that emerges tracks the same logic driving hedging everywhere else: irreversibility and personal-circumstance dependency drive caution, while transactional, lower-consequence queries get more direct treatment.

Life and health insurance sit at the heaviest-hedging end. Both involve irreversible, multi-decade financial commitments, personal health data, and outcomes that depend heavily on individual circumstances like age, health status, and family composition. A query like "do I need a whole life policy at 35" or "which health insurance covers pre-existing conditions" will consistently receive hedged, qualified answers with professional-consult language, regardless of how good the underlying cited source is. Auto and property insurance sit in the middle: more transactional and factual, with AI engines more willing to state specific numbers, particularly when the source is a comparison aggregator or an insurer with published rate data. Travel insurance sits at the lightest-hedging end of the entire category, treated closer to a consumer-product query than a life-planning query, because the purchase moment is short, the stakes are real but lower, and the query is typically more specific and factually answerable.

One piece of real-world evidence supports this pattern directionally, though it is again not Indonesian: the first wave of ChatGPT-integrated insurance products approved by OpenAI, starting with a Spanish provider in February 2026, has skewed heavily toward auto and travel. Liberty Mutual became the first major US carrier to integrate, alongside APRIL Moto and VisitorsCoverage for travel, MoneySuperMarket, GoCompare, Aviva, and Simply Business. None of the first integrations were life or health insurance providers, which is consistent with, though not proof of, the hedging-differential pattern: platforms building AI-native distribution appear to be starting where AI systems already hedge least.

The Four Query Clusters Worth Designing Content Around

Search and AI-prompt behaviour for insurance organises into four recurring clusters, and each earns meaningfully different AI treatment.

Educational and Definitional Queries

"Apa itu asuransi jiwa," "bagaimana cara kerja asuransi kesehatan," "apa perbedaan asuransi syariah dan konvensional." These receive the most complete answers, the fewest disclaimers, and the highest citation willingness of any cluster, because they request education rather than a specific recommendation. This is also the entry point for the roughly 55% of Indonesians the 2025 SNLIK survey found do not yet understand basic insurance concepts, which makes this cluster the fastest place to establish genuine citation authority.

Considered and Comparison Queries

"Asuransi jiwa terbaik untuk keluarga muda," "perbedaan asuransi term life dan whole life," "perbandingan Prudential vs Allianz." AI engines will still name specific products here if the source is authoritative and the claims are properly qualified, but they hedge more than they do on educational queries, and content built for this cluster needs explicit comparison tables and clear qualification language such as "suitability depends on age, income, and health profile," rather than an unqualified verdict.

Urgent and Reactive Queries

"Cara klaim asuransi mobil setelah kecelakaan," "asuransi kesehatan klaim rawat inap prosedur." These seek procedural guidance, and AI engines respond with sequential, how-to-style answers. HowTo schema is the primary citation lever here. Auto claims queries generally receive lighter hedging than health claims queries within this same cluster, consistent with the broader sub-category pattern.

Syariah-Specific Queries

"Asuransi syariah terbaik," "perbedaan takaful dan konvensional." This cluster is structurally underserved rather than simply hedged, for a different reason than the others: authoritative Bahasa Indonesia training data on takaful is thin, which is a content-supply gap rather than a caution-driven hedge, and it is covered in more depth in a companion piece on the sharia insurance content opportunity specifically.

Query Cluster	Hedging Level	Primary Schema
Educational / definitional	Lowest, fullest answers	Article + FAQPage
Considered / comparison	Moderate, still names products if qualified	Product + ItemList
Urgent / reactive (claims)	Low for auto, higher for health	HowTo
Sharia-specific	Underserved rather than hedged	Article + FAQPage, DSN-MUI cited

Where Citations Actually Come From, and a Genuine Caution About This Data

Cross-platform analysis of roughly 38,000 sector-tagged prompts across ChatGPT, Perplexity, Gemini, and Claude found insurance-related answers cite an average of 5.1 sources each, broken down as approximately 27% brand-owned content, 26% review or comparison sites, 18% news or editorial sources, and 5% forum-based user-generated content. This is global, cross-market data, not an Indonesia-specific breakdown, and it should be read that way rather than assumed to describe the Indonesian market precisely.

It is worth being transparent about a genuine data-quality issue encountered while researching this article: one source claimed insurance citations follow an "88% brand-managed pattern," a figure that directly contradicts the more granular 27/26/18/5 breakdown from the same broader research base. The two numbers cannot both be describing the same thing accurately, and rather than pick one arbitrarily, this article uses only the more specific, better-sourced breakdown and discards the vaguer summary figure entirely. This kind of internal contradiction is exactly why cross-validating research claims against each other, rather than citing the first confident-sounding statistic found, matters in a YMYL category where an inaccurate number in published content is a bigger liability than a missing one.

What an Australian Study Says, and Why It Is Not About Indonesia

One widely circulated figure claims "70% of high-intent insurance queries return no brand name at all." Tracing that claim to its actual source matters here: it originates from the Somantra AI Search Visibility Report, a May 2026 study of 34,278 conversations across twenty Australian insurance brands, not an Indonesian study. The same report found that Google AI Overviews and ChatGPT named the same top brand only 27.9% of the time in a given month, up from 23.7% two months earlier. Both figures are genuinely useful as directional context for how volatile and brand-sparse AI insurance answers can be in a comparable, English-language market. Neither should be quoted as a fact about Indonesian insurance search behaviour, because it simply is not what was measured.

This distinction is not pedantic. A content strategy built on the belief that "70% of queries name no brand" specifically in Indonesia would draw a very different conclusion about competitive urgency than a strategy correctly informed that this is an Australian benchmark used only as a directional proxy. Getting the geography of a statistic wrong changes the strategic conclusion it supports.

The Actual Competitive Situation in Indonesia Right Now

Across four independent research passes conducted for this content programme, none found evidence of a disclosed, structured GEO or AEO programme currently run by an Indonesian insurer or broker. Insurer AI investment in Indonesia appears concentrated in claims, underwriting, and service operations rather than answer-engine visibility. That combination, real AI-mediated research demand rising alongside essentially zero deliberate insurer-side GEO activity, is the actual competitive picture, and it means the dominant citation sources in Indonesian insurance answers today are aggregators and OJK's own consumer education content, not insurer-owned material.

YMYL trusted-source pools are documented as unusually sticky once they form, meaning the sources an AI system learns to trust for a sensitive category tend to stay trusted rather than being displaced easily by a later entrant. That combination, of low current competition and high switching cost once a pool solidifies, is the specific argument for treating insurance GEO as time-sensitive rather than something to revisit later once the category matures.

Why Ranking First on Google Is Not a Prerequisite Here

One finding worth building an entire content strategy around: across a large YMYL-focused study, only about 17% of AI Overview citations also ranked in Google's organic top ten, meaning roughly five out of six AI Overview citations came from pages that were not on organic page one at all. For a category entering GEO with little existing organic authority, that is genuinely good news. It means strong extraction-ready structure, answer-first formatting, clean FAQPage schema, genuine HTML comparison tables, can leapfrog pages with higher organic rank, rather than citation being gated behind first achieving organic dominance the slow way.

A second finding from the same research base cuts the other direction and deserves equal weight: roughly 10.4% of AI Overview citations in a large-scale YMYL study were themselves classified as AI-generated content, based on automated detection tools run across nearly 29,000 queries. That is a real quality signal running through the trusted-source pool, and it is also a genuine opportunity: human-reviewed, licensed-professional content is a differentiator precisely because a meaningful share of what AI systems are currently citing was not written or reviewed by a human expert at all. For insurance specifically, where a licensed-reviewer sign-off is close to non-negotiable given the compliance stakes discussed elsewhere in this content series, that differentiation should show up naturally rather than needing to be manufactured.

Realistic timeline expectations matter here too. GEO programmes in finance-adjacent categories typically take four to eight months to show significant citation gains, based on industry benchmarking, which is a materially different cadence from a paid-media campaign and closer to the timeline organic SEO authority-building already requires. Reporting this honestly to a client or an internal stakeholder, rather than promising faster results, is itself part of building the accuracy-first credibility this entire category rewards.

Accuracy Before Visibility

One measurement principle deserves emphasis on its own, separate from every visibility metric discussed above: in a category with real fraud exposure, being cited with wrong information is worse than not being cited at all. If an AI system attributes an incorrect coverage limit or a false exclusion to a named insurer, that is a compliance incident, not a missed opportunity. A brand with high citation frequency and low accuracy is in a genuinely worse position than a brand with moderate citation frequency and high accuracy, which is why any serious GEO measurement stack for insurance tracks Citation Accuracy Rate and Hallucination Rate as the primary, non-negotiable tier, with visibility metrics like Share of Model treated as secondary, exactly the ordering discussed in more operational detail in Arfadia's GEO framework for insurance.

---

Frequently Asked Questions

Does AI hedging on insurance mean GEO does not work for this category?

No. Hedging means the category is treated with YMYL caution, not that AI systems are unwilling to cite anyone. The strategic goal shifts from eliminating the hedge to being the specific source cited alongside it, which is the highest-authority position this category currently offers.

Is the "70% of insurance queries name no brand" statistic true for Indonesia?

That specific figure comes from a May 2026 Australian study of twenty insurance brands, not an Indonesian measurement. No equivalent Indonesian study currently exists, and this figure should only be used as directional, clearly-labelled international context.

Which insurance sub-category should a GEO programme prioritise first?

Educational and definitional content across any sub-category tends to earn citation authority fastest, since it is the least-hedged content type. Within product-specific content, auto and travel insurance currently show lighter AI hedging than life and health, consistent with which insurance lines are being productised first inside AI-native distribution channels internationally.

How reliable are the specific percentage figures cited in AI-citation research generally?

Treat individual vendor-reported percentages with real caution. This article itself found and excluded one internally contradictory figure during research, and separate sources report conflicting AI Overview coverage rates for insurance that could not be reconciled. Directional patterns across multiple independent sources are more trustworthy than any single precise percentage.

What should an insurer do if it is not yet running any GEO activity?

Start with the foundational, educational content tier, since it is both the lowest-hedging and the currently least-contested citation opportunity, and layer in the SEO foundation that AI citation and organic ranking both depend on, rather than treating GEO as a separate initiative built from scratch.

The complete framework for building AI-citation authority in categories where hedging is the norm rather than the exception, drawn from insurance, legal, and healthcare content programmes alike, is covered in Tessar Napitupulu's Cited or Silent. Get the free chapter on YMYL categories at arfadia.com/resources/ebook-cited-or-silent, or talk to Arfadia about building this into a measured programme through our GEO service.

---