Case: three island sites designed as one ecosystem

Portbahn Islay, Bothan Jura Retreat and Isle of Jura could have been three standalone websites. They are built instead as one system with three voices — distinct remits, a shared entity graph, and a written rule set governing which site answers which question. This note explains why, and what the design does in AI retrieval that three separate sites cannot.

Lynton Davidson · Published July 2026 · Last updated 14 July 2026

The problem with one site — and with three

The obvious architectures are both wrong. One site covering accommodation on two islands plus everything a visitor needs to know about Jura would be topically broad, and AI retrieval punishes breadth: systems embed a site's pages and, in effect, average them — pages far from that centre dilute the whole domain's authority on any one subject. Three fully independent sites solve the focus problem and then compete with each other: same owners, overlapping topics, three thin claims to the same questions instead of one strong claim each.

The third option is the interesting one: three sites with deliberately distinct remits, designed together, linking to each other where the visitor's question crosses a boundary. Portbahn Islay answers Islay accommodation and its practicalities. Bothan Jura Retreat answers Jura accommodation in a personal register. Isle of Jura is the destination authority — travel, activities, practical information — and takes the questions the accommodation sites deliberately decline.

The design: one lane per topic

Every topic in the shared territory has one owning site. Ferry practicalities to Islay belong to Portbahn Islay; the Jura passenger ferry belongs to Isle of Jura; what it is like to stay in an off-grid hut belongs to Bothan Jura Retreat. The rule is written down, the allocation of search terms across the three sites is kept in a register, and when any site's content is about to expand we run a cannibalisation audit across the other two first. Where the stronger authority is external — the ferry operator, a distillery — the sites link out and decline to compete at all.

What the sites share is an entity graph, not content. The same places, the same businesses, the same facts, named the same way everywhere — and one pair of human bridges: Pi and Lynton, the hosts on every site. When a retrieval system follows a multi-hop question — stay on Islay, day trip to Jura, who can tell me how? — the relationship between the sites is stated in prose on each of them, not left for the system to infer from links.

Why it compounds in AI retrieval

AI search does not run your question; it plans research around it. A single prompt fans out into ten to thirty synthetic sub-queries — variants, adjacent questions, verification probes. Published fan-out research puts the citation pattern plainly: a domain appearing once or twice across those sub-queries is a marginal candidate; a domain family appearing across many of them becomes hard to leave out of the answer.

One broad site cannot occupy many branches of a fan-out — its pages are averaged into one thin claim. Three focused sites can each win their own branches: the accommodation question, the ferry question, the what-to-do question, the which-island question. Each site's win strengthens the others, because the entity graph ties them together and the answer engine keeps meeting the same consistent facts from three directions. That is the design bet: an ecosystem can saturate a fan-out; a single site can only rank in it.

The discipline that makes it work

• No duplication, ever. Repeated content is the failure mode that turns an ecosystem back into three competing sites. Shared facts live in structured entity records; each site renders its own editorial.
• Term allocation is a register, not a habit. Which site owns which query family is recorded, and disputes are settled before writing, not after publication.
• Cannibalisation audits are mechanical. A script compares the sites' query sets before any expansion; overlap is surfaced as a report, not discovered as a ranking accident months later.
• Authority decisions stay human. Which site a piece of content belongs to — or whether it should exist at all — is an ownership decision, made by the owners.

Where it stands

Portbahn Islay is rebuilt and live with the full architecture — entity records, canonical content blocks, structured data end to end. Bothan Jura Retreat's content and data layer are largely complete and its build is in progress. Isle of Jura runs on its existing platform with migration planned. The architecture is ahead of the deployment, deliberately: the rule set and the measurement system came first, so each site lands into a frame that already knows what it owns.

The compounding effect is the hypothesis the measurement system is built to test. Weekly probes ask live AI assistants the questions our guests ask and record which sites are cited; as the second and third sites land, the same probes will show whether the ecosystem outperforms the sum of its parts. We expect it to. We will know, either way.

The general pattern

Nothing here is specific to islands. The architecture applies wherever one organisation speaks to distinct audiences whose questions overlap: a destination and the businesses within it; a brand and its community; an institution's public site and its student-facing one. The design questions are always the same — what does each site own, what do they share, where do they defer, and how will you know it is working. This is the architecture the studio brings to client work; the islands are where it is proven first.