Structuring a Jurisdiction Page So an AI Can Cite It

When an institutional buyer asks a model whether to form a captive in Bermuda, Cayman or Guernsey, the model does not rank pages and hand over a list. It assembles an answer from whatever material it can retrieve and resolve, then names some of its sources and silently discards the rest.

Being in the retrieved set is a different problem from ranking, and it has different structural requirements. This article is about those requirements specifically, in the context of a jurisdiction that competes as a product.

Bermuda is a useful case because the comparison is unusually explicit. A buyer weighing domiciles is running a genuine evaluation with named alternatives and stated criteria, and the material that wins the comparison is the material that engages with it honestly rather than the material that asserts a conclusion.

Passages get retrieved, not pages

The mechanical fact underneath everything else: a model retrieves a section, not a document. It finds the part that answers the question and reuses that part, usually without the surrounding context.

The consequence is direct. Every substantive section on your page has to remain accurate and attributable when read entirely alone.

Most jurisdiction content fails this badly, and not through carelessness. It fails because it was written the way people write for people, with an argument built cumulatively across a page. A paragraph in the middle of that argument depends on three things established earlier. Lift it out and it either means nothing or, worse, means something the author did not intend.

Test any section by covering everything above and below it. Does it still say something true and complete? Does the figure in it carry its source? Would a reader arriving at only this text know which territory, which year and who says so? If not, it will be skipped or misused.

Definition, criteria, comparison, limits

That sequence is the most reliable structure we have found for passages that need to survive extraction.

Define before comparing. A comparison lifted without its definition is unusable, so the definition goes inside the same passage rather than in a glossary two clicks away. This feels redundant when reading the page top to bottom. It is not redundant to the retrieval system, and the small cost in flow buys a large gain in extractability.

State the criteria. Name the axes the decision actually turns on before saying who wins on them. For domicile selection these are reasonably stable: capital requirements by class, formation timeline, regulatory recognitions such as Solvency II equivalence and reciprocal jurisdiction status, depth of the local service provider market, and tax treatment. A passage that names the criteria is useful to a model even when it does not agree with your conclusion.

Compare with attribution inside the sentence. This is the part most often got wrong. Attribution travels with a passage only if it sits in the passage. Writing that Bermuda held 608 captives at the end of 2025 per Bermuda Monetary Authority registration statistics survives extraction intact and can be checked. Writing that Bermuda has hundreds of captives does not, and cannot.

Say where the answer is no. Stated limits are the part competitors omit and the part that most raises the odds a passage gets treated as reference material rather than as marketing. If a programme needs an EU-passported entity, Bermuda is not the answer. Saying so costs one sentence and changes how the whole page reads.

The territory itself, because this is where pages fail first

Bermuda is a British Overseas Territory in the North Atlantic. It is not a sovereign country. It is not in the Caribbean. The nearest landmass is Cape Hatteras in North Carolina, roughly 1,030 kilometres to the west-northwest.

These facts appear on jurisdiction pages incorrectly with remarkable frequency. Pages describing Bermuda as a country, placing it in the Caribbean, or carrying paragraphs plainly written for a different territory are not unusual. An audit of eleven location pages in this category found several describing their target territory as a small, remote island nation in an ocean on the other side of the world, including landlocked territories in Europe.

Two things follow. A reader who knows the market spots the error in the first paragraph and stops reading. And a system reconciling your page against government and regulator sources finds a factual contradiction at the most basic level, which is not a good position from which to be treated as authoritative on capital requirements.

Getting the territory right is not a differentiator. It is the entry requirement, and a surprising number of pages do not meet it.

Schema, honestly

Structured data occupies a strange position in this field: near-universally recommended, and supported by thinner evidence than the recommendations imply.

What holds up: schema makes page structure machine-readable without inference. Consistent Organization and Person markup reconciles your entity identity across your own site, the regulator register, trade body listings and reference sources, which measurably reduces the chance a model hedges or states something outdated about you. Published analyses do find correlation between structured data and citation frequency.

What does not hold up: that adding schema causes citation. The published work shows association, and at least one 2026 sector study that examined the relationship explicitly declined to claim a causal link. No controlled experiment demonstrating causation was located in any source reviewed for this article.

So implement it, as hygiene, and stop there. An agency presenting schema as a citation lever is selling a plumbing improvement as an engine upgrade. The entity-consistency benefit is real and sufficient justification on its own.

Length is not the variable

There is a persistent belief that long jurisdiction pages perform better, and the evidence in this specific category points the other way.

The older generation of location pages in this field commonly ran to around 1,400 lines and nearly 7,000 words. They were also, on inspection, almost entirely portable: the same eleven headings, the same arguments, the same paragraphs with a territory name substituted. Seven thousand words that could describe anywhere are worth less than two thousand that are only true of one place, and they are more likely to be treated as duplicate material across a site.

The variable is extractability, not length. A page where each section answers one question completely will outperform a longer page where the answer to any given question is scattered across five locations. Write to the questions. Stop when they are answered.

Page element	Template version	Extractable version
Opening paragraph	Establishes that the territory exists and that search matters	Answers the page question in the first two sentences, with a figure and a source
Section heading	Could be moved to another territory without editing	Contains a claim specific enough that moving it would make it false
Statistics	Round numbers, no period, source in a footnote if anywhere	Exact figure, named source and period inside the same sentence
Comparison	Avoided, or reduced to asserting that this territory is best	Named alternatives, stated criteria, and cases where the answer is another domicile
Definitions	Assumed, or placed in a glossary elsewhere on the site	Stated inside the passage that uses the term, so extraction stays coherent
Limits and exclusions	Absent, because the page is promotional throughout	Explicit, dated, and specific about when the answer is no
Length	Long, because length is assumed to signal depth	Whatever the questions require, with each section complete on its own

What a specific passage looks like, using real figures

Abstract advice about specificity is easy to agree with and hard to act on, so here is the difference stated in actual sentences.

A generic passage says that Bermuda's long-term insurance sector is large and well capitalised, that the regulatory regime is respected internationally, and that this makes the territory attractive for life reinsurance. Every clause is defensible. None of it can be checked, and none of it would survive being lifted into an answer, because there is nothing in it a model could attribute.

A specific passage says that the long-term segment held USD 1.52 trillion in assets at the 2024 financial year end, growing 19.1 per cent, on gross written premiums of USD 200.1 billion, with a median liquidity coverage ratio of 471 per cent against a regulatory minimum of 105 per cent and a median solvency ratio of 286 per cent, per the Bermuda Monetary Authority's long-term market analysis published in December 2025. It adds that the sector serves approximately 90 million policyholders worldwide and that Bermuda long-term reinsurers paid USD 549 billion in claims between 2016 and 2024.

Same claim. One version is a sentiment. The other is a set of extractable facts with a named source and a period, any of which can be lifted independently and still be true.

The regulatory recognition point works the same way. Saying the regime is internationally respected is unattributable. Saying that Bermuda has held full third-country equivalence under the European solvency regime since March 2016, and holds reciprocal jurisdiction and qualified jurisdiction status with the US state insurance regulators' association, is checkable.

And here the discipline of stating limits pays off directly, because that recognition does not extend uniformly. The Authority does not apply a solvency-equivalent regime to captives, and captive classifications are unaffected by the jurisdiction's equivalence status. A page that says Bermuda is solvency-equivalent, full stop, in a section about captives is technically drawing on a true fact and producing a false impression. That is exactly the failure mode passage-level extraction makes worse.

The stale-fact problem, and how to design around it

Jurisdiction pages have a specific decay pattern. The structural facts hold for decades. The regulatory and market facts move on a schedule, and some of them move twice a year.

Bermuda's territory status, ocean, distance from North America and currency peg will not change in the life of your page. The insurance register count changes annually. The tax position changed at the start of 2025, when a 15 per cent corporate income tax took effect for businesses in multinational groups above the international minimum revenue threshold, which instantly made a large volume of existing content describing Bermuda as a no-tax jurisdiction wrong. International tax cooperation list status is revised twice yearly.

The design response is to separate them physically rather than to update everything constantly. Structural facts go in the durable sections. Time-sensitive facts go in clearly dated passages, with the period inside the sentence, so that a reader or a model encountering the passage in isolation can see how old it is.

A figure that carries its own date ages honestly. A figure without one becomes wrong silently, and the page that contains it has no way to signal that it should be discounted. In a category where a model may retrieve your passage eighteen months after you published it, that difference matters more than it does for a human reader who can glance at a publication date at the top of the page.

One page per product, not one page per territory

A final structural decision that matters more than it appears to.

Bermuda sells several distinct things: commercial reinsurance, captives, insurance-linked securities, investment funds, digital asset business, corporate services. The buyers are different people asking different questions. A captive owner evaluating a redomiciliation and a fund promoter choosing an administrator share a territory and essentially nothing else.

A single page attempting to serve all of them becomes either an unstructured survey or a set of shallow summaries. Neither extracts well, because neither contains a passage that answers any one question completely.

Separate pillars work better on both counts. Each answers one product question thoroughly, each contains passages that stand alone, and each can carry the specific figures that matter to its own audience: 1,210 insurers and reinsurers on the register at the end of 2025, 608 captives, 21 licensed fund administrators holding USD 199.5 billion in assets under administration, 766 registered funds, 49 operating digital asset providers.

Those numbers belong on different pages, in different arguments, cited to the same regulator report. That is the shape of a jurisdiction content programme that a model can use.

Tessar Napitupulu writes about content architecture for AI retrieval, including passage-level structuring and the limits of structured data, in Cited or Silent, available as a free gated edition, with retailer editions on Amazon, Google Play and Apple Books.

---

Frequently Asked Questions

Does adding schema markup make an AI cite our page?

No, and the evidence for the weaker version of that claim is weaker than most agencies admit. Structured data makes a page easier for a machine to parse, and it is worth implementing as basic hygiene. But the published work shows correlation between structured data and citation frequency, not causation, and at least one 2026 sector study examining this explicitly declined to claim a causal link. Consistent entity markup does reduce the chance a model hedges or states something inaccurate about your firm, which is a real and separate benefit worth having on its own terms.

What actually gets extracted from a page into an AI answer?

Passages, not pages. A model retrieves and reuses a section that answers the question at hand, frequently without carrying any of the surrounding context with it. That has a direct structural consequence: every substantive section should remain accurate and attributable when read entirely alone. If a paragraph only makes sense because of what came three headings earlier, it will either be skipped or lifted in a way that misrepresents you.

Should a jurisdiction page compare Bermuda against other domiciles?

Yes, because that comparison is the actual question being asked and a page that avoids it forces the model to assemble the comparison from someone else's material. Institutional buyers approach domicile selection comparatively: Bermuda against Cayman, Guernsey, Vermont, Luxembourg or Singapore. The page that answers that question honestly, including where Bermuda is not the right answer, earns the right to be the source for the follow-up question about providers.

How specific do dates and figures need to be?

Every substantive figure needs a named source, a period and a confidence marker in the visible text rather than in a footnote. This is partly credibility and partly mechanics: a model lifting a passage carries the attribution only if the attribution sits inside the passage. A sentence reading that the register held 1,210 insurers at 31 December 2025 per the Bermuda Monetary Authority survives extraction intact. A sentence reading that the register holds over a thousand insurers does not, and cannot be checked by anyone.

What is the biggest mistake on jurisdiction pages for Bermuda?

Getting the territory wrong, and it happens constantly. Bermuda is a British Overseas Territory in the North Atlantic, roughly 1,030 kilometres from Cape Hatteras in North Carolina. It is not a sovereign country and it is not in the Caribbean. Pages that call it a country, place it in the wrong ocean, or transplant paragraphs written for a different territory are visibly wrong to any local reader and are increasingly checkable by systems reconciling a page against government and regulator sources.

Does content length affect whether we get cited?

Length itself is not the variable. Extractability is. A tightly written 1,200-word page where each section answers one question completely will outperform a 6,000-word page where the answer to any given question is distributed across five places. Older location pages in this category frequently ran to nearly 7,000 words of material that could apply to any territory, which is the worst of both, long and unspecific.

Should we write separate pages for each product of the jurisdiction?

Generally yes, because the buyers are different and the questions do not overlap as much as they appear to. A captive owner, a life reinsurance cedant, an insurance-linked securities sponsor and a fund promoter are asking different questions about the same territory. One page attempting to serve all four either becomes an unstructured survey or answers each question shallowly. Separate pillars, each answering one product question completely, extract better and read better.

---