JJ MARINE

JJ Marine sells and charters yachts in Dubai off a Wix site with 18 live URLs. The catalogue lived as PDF brochures, which meant the specifications buyers search for existed nowhere a crawler or an answer engine could read them, and the inventory could only be updated by hand, page by page. A dealership whose stock turns over cannot run on hand-edited pages. It also sells into a market where the buyer is as likely to read Arabic or Russian as English, and the old site was English only.

Rebuild the site as a database-backed catalogue where the boat row is the source of truth, then treat search visibility as a first-class output of the data model rather than a marketing layer bolted on afterwards. Every boat renders its specifications as on-page HTML and emits Product plus Offer structured data from the same row. The migration off Wix was planned before any of it: all 18 live Wix URLs accounted for — 6 kept their slug, 7 got a 308, 4 became hard 410s — plus 7 redirects for the old brochure PDF deep-links, so nothing that had accrued authority landed on a soft 404. The build itself ran as agent work under a coordination protocol, with adversarial review agents as a scheduled phase rather than a nice-to-have.

The plan was 7 epics and about 50 micro-sprints. It finished at 111 sprint commits, numbered up to S165, across 22 epics. Epics 7 through 21 were all emergent, driven by feedback while the thing was already running: a matcher for buyers who do not know what they want, mobile touch behaviour, four languages, a launch runbook, a self-hosted analytics layer, an admin worklist. The gap between the plan and the log is not a planning failure so much as the actual shape of building in-house, where the people who own the business see the site every day and keep having opinions. It is worth stating plainly because a 50-sprint estimate is what I would have committed to.

What it replaced was a Wix site with 18 live URLs and a catalogue delivered as PDF brochures. The rebuild moved 46 boats (27 power, 12 sail, 3 pre-owned, 4 charter) into a Supabase table where every row auto-generates its own page plus JSON-LD, added 14 editorial guides, and grew the sitemap from 80 URLs to 320 as the four locales came online. Migration discipline was written before the launch, not after: 15 redirect rules plus 4 hard 410s, covering every one of the 18 old URLs that did not keep its slug.

Worth being precise about, because the phrase "AI" attaches itself to marine dealerships easily: the public site has no chatbot, no LLM in the shipped bundle and no AI dependency in its package.json. What runs internally is a separate matter and stays internal. Everything described on this page is the public site.

The true story is better. The site is built by agents, under a protocol. S156 wired the repo into an Aethereum coordination room with 13 agent surfaces and autopilot on, and AGENTS.md makes the loop mandatory: pull team context before touching cross-service code, share intent, declare a contract for every new interface, claim and release before a refactor, check the blast radius before changing a contract others depend on. Agent-to-agent briefs exist as committed artifacts. The whole of S158 is one agent writing a brief for another about the leads view, ending with a question back it refuses to build past. S162's commit body opens by answering it. Adversarial review agents are a phase in the build: S140 fixed three live search defects found in S138's own code, and S162's three headline bugs came out of adversarial review rather than from a user.

It also collided. MASTERPLAN records that sprint numbering is broken for 2026-08-05, because two agents worked the repo concurrently and S156, S157, S158, S159 and S160 each name two different commits. S162's commit body records two renames being swept into the other agent's commit, because git add -A from either working tree picks up whatever the other has staged. Nobody lost work, but the history is no longer readable by sprint number and has to be disambiguated by hash. That is a real cost of a genuinely new way of working, and it is in the repo rather than tidied away.

Next.js 16.2.9 App Router on Turbopack, React 19.2.4, TypeScript strict, Tailwind CSS v4 with @theme tokens, Motion 12.42 under LazyMotion, next-intl 4.13, Supabase, Vercel. 24 page routes. The accent colour is a single token, so a rebrand is a one-line change rather than a search-and-replace across components.

The catalogue is the spine. Boats live in a Supabase table and every row generates its detail page and its Product plus Offer JSON-LD from the same source, alongside filter, sort, facet and full-text search over the collection. Specifications render as on-page HTML instead of living only inside a brochure PDF, which was the single biggest search and answer-engine win in the whole build, and 44 of the 46 boats still carry their brochure as a download. dynamicParams is false on /boats/[slug], so a slug that is not in the table returns a true 404 rather than a rendered empty shell; boats that do exist revalidate hourly through ISR. 41 of 46 have multi-image galleries.

Internationalisation runs across en, ar-AE, ru and fr with localePrefix 'as-needed' and full RTL for Arabic. S143 to S147 alone moved roughly 500 strings, on top of 4 metadata catalogs at 146 keys each, 138 boat translations and 42 translated guides. That is what turned an 80-URL English site into a 320-URL sitemap with an hreflang matrix and x-default across all four.

The discovery layer is deliberately wide: JSON-LD for AutoDealer, CollectionPage, Product with Offer, BreadcrumbList, WebSite, BlogPosting, Person, Service and FAQPage, an Atom feed, llms.txt and llms-full.txt, an AI-crawler group in robots.txt naming 12 bots (GPTBot, ClaudeBot, meta-externalagent and others), plus an IndexNow key file and ping script. Lead capture writes to an insert-only Supabase table behind RLS, with Turnstile and a honeypot live and Upstash rate limiting wired but fail-open until its keys are set in production. Analytics is self-hosted: a site_events table and a navigator.sendBeacon client, with no third-party analytics as a dependency (Vercel's own Analytics and Speed Insights tags stay mounted as a free cross-check, but nothing reads from them). A password-gated /admin, unlinked, disallowed in robots.txt and excluded from the sitemap, carries a leads worklist, lead volume over time, traffic, boats ranked by enquiries, the newest 50 leads and CSV export. A separate social pipeline renders 20 posts and 92 slides from live catalogue data through satori and resvg and pushes approved ones to Buffer as drafts.