68%
Bermuda insurance groups already using AI or machine learning systems, against 38% across the surveyed market and 23.7% among property and casualty insurers
Source: BMA, Bermuda Insurance Sector AI and Machine Learning Survey, 2022 Report (VERIFIED)
~50%
Bermuda residents using AI for personal purposes at least weekly, in a non-probability online panel of 248 adults
Source: Narrative Research Bermuda, fielded Dec 2025 to Jan 2026, via Bernews Jan 2026 (REPORTED, no margin of error can be ascribed)
~11%
Overlap between the domains ChatGPT and Perplexity cite for the same query, with around half of cited domains rotating month to month
Source: Similarweb citation analysis, Jan to Feb 2026 (REPORTED)
2023
The year Arfadia began documenting its GEO practice, before the category had a common name
Source: Arfadia dated archives
About GEO & AEO for Bermuda
Bermuda does not sell a product so much as a jurisdiction. When a model is asked to compare domiciles, it assembles that answer from whichever sources it can resolve and trust. Getting into that set is a different job from ranking.
The Buyers Are Institutional, and They Are Not On The Island
Bermuda is a British Overseas Territory in the North Atlantic with roughly 64,000 residents, and that number is close to irrelevant to this service. The commercial audience is the risk manager, broker, cedant, captive owner, fund counsel and institutional allocator deciding where to place a programme. More than 70 per cent of Bermuda long-term insurers' reserves originate in the United States, with Asia second and European business including the UK below 5 per cent. Those people research domiciles the way everyone now researches anything technical: they ask a model to compare the options before they ask a colleague. There is no published measurement of how often AI systems are consulted on domicile selection specifically, and we have not found one. What is measurable is the surrounding behaviour, and it points one way.
Two Adoption Numbers, Two Different Questions
The most useful AI adoption figure for this market comes from the regulator rather than a vendor. The BMA surveyed its own insurance sector and found 38 per cent of respondents overall using AI or machine learning, but 68 per cent among insurance groups, a split the Authority attributed to the scale of their international operations. Among the 62 per cent not yet using it, 23 per cent planned to adopt within five years. Separately, and measuring something entirely different, a Bermuda online panel of 248 adult residents found around half using AI weekly for personal purposes and close to six in ten employed respondents using it weekly at work. These two figures answer different questions about different populations, one institutional and one residential, and merging them would produce a number that describes nobody. The residential figure is also a non-probability online panel, so no margin of error can properly be attached to it.
One Domicile Question, Two Different Source Sets
Ask two engines the same domicile comparison and they will not reach for the same material. The answer a buyer reads depends on which engine they happened to open.
Should we form our captive in Bermuda, Cayman or Guernsey, and what changes after the 15% corporate income tax?
Tends to resolve against durable reference and regulator material:
Regulator and government publications with stable URLs
Encyclopaedic and reference entries on the jurisdiction
Regulator statistics
Reference entries
Statute and rules
Should we form our captive in Bermuda, Cayman or Guernsey, and what changes after the 15% corporate income tax?
Tends to weight recent, dated third-party commentary:
Professional advisory briefings published this quarter
Trade press covering registration and premium statistics
Advisory briefings
Trade press
Recent commentary
~11% overlap
Analysis of roughly 600,000 citation events found only about 11 per cent overlap between the domains ChatGPT and Perplexity cite, with around half of cited domains changing month to month. Optimising for one engine leaves most of the answer surface uncovered, and a single audit goes stale within a quarter.
Illustrative of the source-divergence pattern documented in the cross-validated research behind this page. It is not a captured screenshot of a specific query, and the source categories are inferred from citation patterns rather than measured for Bermuda queries specifically. No study of AI citation shares for Bermuda domicile prompts was found in any of the eight research documents reviewed.
Why Citation, Not Ranking, Is the Metric Here
Ranking is a position in a list of links. Citation is whether a model names you inside the answer, usually without a click. In financial content the two barely overlap, which is why treating GEO as a phase of SEO produces a measurement gap rather than a saving.
AI answer citations in finance, versus organic ranking
Finance citations also ranking in organic top tenlowest overlap of any tracked industry
Finance citations from pages outside the top hundredinvisible to conventional rank tracking
Comparable overlap across YMYL categories generallyhealth, finance, legal
MADX Digital analysis (REPORTED, single vendor source, global rather than Bermuda-specific). We report it because it is directly relevant and we have found no better measurement, not because it is definitive.
30-40%
Visibility Lift From Structural Changes, Peer Reviewed
The GEO study presented at ACM SIGKDD 2024 by researchers from Princeton and IIT Delhi found that adding statistics, cited quotations and direct-answer structuring raised source visibility by roughly 30 to 40 per cent in its benchmark. Independent and peer reviewed, unlike most figures circulating in this category. Its external validity for today's commercial models remains unproven, and we say so.
No
Schema Does Not Guarantee Citation
Structured data makes a page easier for a machine to parse, and we implement it as hygiene. But the published evidence shows correlation, not causation, and no study demonstrates that adding schema causes a model to cite you. Any agency presenting schema as a citation lever is overselling it.
AI answer prevalence varies enormously by query category and by how each tracker defines a trigger, with published estimates spanning single digits to well above half. This page deliberately states no single prevalence figure, and no Bermuda-specific measurement was available in any source reviewed as of August 2026.
Our GEO & AEO Services for Bermuda
Becoming the source a model reaches for when an institutional buyer asks about the jurisdiction, and proving it with dated evidence rather than a promise.
Jurisdiction Answer Architecture
Extractable pages for each product of the domicile, captives, reinsurance, insurance-linked securities, funds and digital assets, built definition first, then criteria, then comparison, with scope limits and dates stated in the text rather than in a footnote.
Models cite passages, not pages, so each section is written to survive being lifted out of context and still be accurate.
Entity Resolution and Authority Signals
Your legal entity name reconciled across your own site, the BMA register, trade body listings, professional directories and reference sources, with Organization and Person schema and named, credentialed reviewers.
A model has to work out who you are before it can recommend you. In regulated finance that identity graph does more work than any keyword list.
Third-Party Authority Building
Earned presence in the reinsurance, captive and fund trade press and the professional commentary layer around it, because those are the sources both institutional readers and models reach for on Bermuda questions.
Two credible studies disagree on how much of AI visibility comes from owned versus third-party sources, so we build both rather than betting the engagement on one number.
Multi-Engine Monitoring With PromptWatch
A versioned panel of institutional prompts tested across multiple engines on a fixed schedule, with the query, the date, the engine and the answer captured every time.
Since roughly half of cited domains rotate month to month and cross-engine overlap runs near 11 per cent, a single audit is a snapshot. Monitoring is the deliverable.
Privacy-Safe Measurement Design
The monitoring layer is designed to hold no personal information at all: public queries, corporate entity names and URLs only. The Privacy Commissioner's guidance is explicit that anonymised information falls outside the overseas transfer restrictions.
Where personal information genuinely is involved, a Section 15 processing agreement and a documented data-flow map come with it.
Regulated Claim Review
Every performance and comparative claim checked against the conduct framework that actually applies to your licence class, which for a wholesale reinsurer or captive is not the framework that applies to a domestic retail insurer.
Runs alongside
our companion SEO service for Bermuda rather than replacing it.
in a domicile comparison, the cited source sets the shortlist
Why Choose Us as Your GEO Agency for Bermuda?
Documented GEO Practice Since 2023, and a Written Refusal To Invent Numbers
This is a category where almost every published figure comes from a vendor selling the service. For a BMA-regulated buyer whose compliance function will read the proposal, an agency that separates what is measured from what is inferred is worth more than one with a bigger headline number.
3
Offices: Jakarta, Bandung, Bali
Anchored To Peer Review, Not Vendor Statistics
Our method rests on the GEO study presented at ACM SIGKDD 2024, which found roughly a 30 to 40 per cent visibility lift from statistics, cited quotations and direct-answer structuring. We name the study, the venue and its limits, and we leave out the circulating vendor multiples we cannot verify.
We Report What Is Measurable, Separately
Citation presence, share of visible answers across a dated prompt panel, and accuracy of what models say about you. Pipeline influence is reported apart from all of that, with the attribution method stated. We will not multiply a citation by an average mandate value to produce a return figure, because that link is not measurable.
Monitoring Built To Hold No Personal Data
The Privacy Commissioner's guidance confirms anonymised information sits outside the overseas transfer restrictions. We design the measurement layer to stay there by construction, so the reporting you buy does not create a Section 15 question you did not previously have.
Every Engine, Every Cycle
Cross-engine citation overlap runs near 11 per cent and roughly half of cited domains rotate monthly. We test a versioned prompt panel across engines on a fixed schedule and keep the raw captures, because a one-off audit in this category expires faster than the invoice.
Explore Related Services
GEO for a Bermuda platform works hardest when the rest of the stack is pointed the same way.
Want To See What A Model Says About Your Domicile Today?
We will run a dated prompt panel across multiple engines on the domicile questions your buyers actually ask, and show you the queries, the dates and the raw answers, including where a model gets you wrong. Contact our team to get started.
Request Your Free GEO Audit
Frequently Asked Questions About GEO & AEO in Bermuda
What is the difference between GEO, AEO and SEO, and does a Bermuda platform need all three?
SEO works on where you appear in a list of links. AEO works on being the extracted answer to a specific question. GEO works on being named as a source inside a generated answer, often with no click at all. The three share a technical and content foundation, which is why they are usually sold together, but they are measured differently and can move in opposite directions. A page can rank first and never be cited, and a page well outside the first results can be cited repeatedly. For a Bermuda platform selling into institutional buyers, the citation layer matters disproportionately because the buying question is comparative and the comparison increasingly happens inside a generated answer before any vendor is contacted.
Is a GEO agency just an SEO agency that added AI to its service list?
Often, yes, and it is a fair question to ask any provider including us. The honest test is not whether an agency uses the word GEO but whether it can show you a dated prompt, the engine it was run on, the raw answer returned, and what changed after intervention. We began documenting our GEO practice in 2023, before the category had a settled name, and we will run a live panel on your domicile questions before you sign anything. Ask for the same from anyone else you are considering. An agency that cannot produce a dated capture is selling the label.
Do AI systems actually influence domicile selection, and can you prove it?
We cannot prove it, and no source we reviewed can either. There is no published measurement of how often AI systems are consulted specifically on captive or reinsurance domicile decisions, and none of the eight research documents behind this page contained one. What is measurable is the surrounding behaviour. The BMA's own survey of its insurance sector found 38 per cent of respondents using AI or machine learning, rising to 68 per cent among insurance groups, with a further 23 per cent of non-adopters planning adoption within five years. A separate residential panel found around half of Bermuda adults using AI weekly. Those are adjacent indicators, not proof of a purchasing mechanism, and we present them as such rather than converting them into a claim about your pipeline.
Which engines do you monitor, and why not just optimise for the biggest one?
Because the engines do not agree with each other. Analysis of roughly 600,000 citation events found only around 11 per cent overlap between the domains ChatGPT and Perplexity cite, and about half of cited domains changed month to month. Optimising for one engine therefore leaves most of the answer surface uncovered, and a single point-in-time audit expires within a quarter. We baseline and re-test across ChatGPT, Google AI Overviews and AI Mode, Perplexity, Gemini, Copilot and Claude, keeping the raw captures so a change in visibility can be distinguished from normal engine churn.
How do you measure GEO success when a single mandate can be worth millions?
By reporting what is measurable and refusing to manufacture the rest. The primary measures are citation presence across a versioned and dated prompt panel, share of visible answers within that panel, factual accuracy of what models say about your entity, and non-brand impressions segmented by searcher country. Qualified enquiries are reported with the source captured explicitly. Pipeline influence is reported separately and only with the attribution method stated in the same breath. What we will not do is multiply a citation by an average mandate value to produce a return figure, because the causal path from an AI citation to a treaty placement or a fund mandate is not measurable, and presenting it as though it were would be dishonest to a buyer whose compliance function is reading the same document.
Will structured data or schema get us cited?
It helps a machine parse the page, and we implement it as basic hygiene, but nobody has shown that it causes citation. The published work in this area demonstrates correlation between structured data and citation frequency, not causation, and at least one 2026 sector study explicitly declined to claim a causal link. Consistent entity information does reduce the chance a model hedges or states something inaccurate about you, which is a real and separate benefit. Any agency presenting schema as a citation lever is overselling a hygiene measure.
We are BMA-regulated. Will AI-optimised content create a conduct problem?
It can, and the answer depends on your licence class rather than on a single universal rule. The Insurance Code of Conduct, revised in August 2022 under the Insurance Act 1978, addresses communications being fair, clear and not misleading, and advertisements, within its conduct-of-business section. That framework was deliberately split between retail and non-retail arenas, on the reasoning that the imbalances in markets serving individuals and small businesses do not apply in the wholesale sector. Investment business and banking sit under separate codes with their own advertising definitions. So the obligations attaching to a domestic retail insurer are not automatically those attaching to a wholesale reinsurer, a captive or a fund administrator. We establish which framework governs your entity before drafting, and every comparative or performance claim is reviewed against it.
Does prompt and citation monitoring create a PIPA transfer problem?
Designed properly, no, and this is worth understanding rather than taking on trust. The Personal Information Protection Act 2016 came fully into force on 1 January 2025, and Section 15 governs transfers of personal information to an overseas third party. The Privacy Commissioner's guidance states that where information has been anonymised, those restrictions do not apply and it may be transferred outside Bermuda freely. Prompt panels, engine responses, corporate entity names and URLs contain no personal information, so the monitoring layer can be built to sit outside the transfer regime entirely. Exposure appears when lead forms, CRM records, session-level analytics or monitoring of named individuals enter the scope, and for those we bring a Section 15 processing agreement and a documented data-flow map. Your organisation remains responsible under the Act either way, which is exactly why we would rather the reporting hold no personal data at all.
What happens if a model says something inaccurate about our entity?
This is one of the more common reasons a regulated Bermuda platform comes to us, and it is a repair job rather than a promotion job. Models reconcile an entity against whatever sources they can resolve, so inaccuracies usually trace to inconsistent legal entity naming across your site, registers, directories and reference sources, or to outdated third-party material that is easier to retrieve than your current position. We document the specific inaccuracy with the engine, query and date, identify the sources likely producing it, correct what is within your control, and pursue correction of the third-party material where that is possible. We cannot compel a model to change its output, and no agency can. What we can do is remove the material the model is reasoning from and monitor whether the answer changes.
Can you guarantee we will be cited in ChatGPT or Google AI Overviews?
No, and any agency offering that guarantee is describing something it does not control. The engines control retrieval and generation, their source preferences shift, and roughly half the domains they cite change from one month to the next. Nor is citation purchasable: the platform operators state that product results in their generated answers are selected independently of advertising or partnership arrangements, with sponsored placements labelled separately. What we commit to is deliverables and method, a versioned prompt panel, a fixed testing cadence, raw dated captures, correction of inaccuracies within your control, and structural work grounded in peer-reviewed findings rather than vendor claims. Meaningful movement typically needs at least 60 to 90 days because models have to recrawl and reprocess sources, which is why we do not scope this work in shorter engagements.
ChatGPT source pattern
Perplexity source pattern
