The Structured Data Asset New Caledonia Already Has

Ask a French-language AI assistant for the best GEO agency in New Caledonia and you get metropolitan French hosting companies, plus a scattering of local businesses that happen to contain the word agence and turn out to sell real estate or surveying services. Ask instead about AI visibility for a company in New Caledonia and a genuinely Caledonian page comes back.

We ran those two prompts in August 2026. It was a small probe through one web-search layer, not a geolocated session in Noumea and not a controlled benchmark, and we present it with those limits attached because three prompts do not establish how AI systems treat an entire territory. What the probe suggests is conditional rather than absolute: where explicit local content exists, it can be retrieved; where the local entity picture is thin or ambiguous, metropolitan sources fill the gap.

The gap-filling is the mechanism worth understanding. A retrieval system does not return nothing when it lacks local sources. It returns the nearest available thing, which for a French-speaking territory means France. Every metropolitan page that surfaces in place of a Caledonian one is a slot that a local business could have occupied by being findable and unambiguous.

The unusual part is that this territory already owns most of the infrastructure needed to fix that. Most small markets do not.

Apidae, and what 3,200 records actually represent

New Caledonia has run on Apidae since 2019. It is the shared structured tourism database used across France and its overseas territories, with more than 25,800 users across the network.

Locally, Sud Tourisme Nouvelle-Caledonie keeps roughly 3,200 destination records current on a daily basis, maintained through the network of tourist office advisers. Those records feed the destination's own website, its events listings, its print and digital promotion, and a number of partner sites including Nouvelle-Caledonie Tourisme.

Read that as an infrastructure fact rather than a tourism administration detail. A territory of 264,596 people maintains a daily-updated, structured, machine-readable record of who provides what and where, consumed by multiple downstream publishers. That is precisely the kind of consistent institutional source that answer engines favour, and it exists here already, funded and staffed, whether or not any individual business is paying attention to it.

The platform has also been moving in the same direction as the rest of the field. Apidae added a generative AI feature for producing record descriptions in October 2024. That cuts both ways for a business owner. It lowers the effort of having a description at all, and it raises the risk that your record ends up reading like everybody else's, which is the opposite of what makes a source worth citing.

The checks worth running on your own record

Six questions, each answerable in an afternoon, and each one addressing a specific way records fail to earn citation.

Is the record present, and is it current today rather than in 2023? A great deal changed in this territory after May 2024. Opening hours, service lines and even trading status moved. A record that was accurate two years ago and has not been touched since is worse than absent, because it contradicts your website and gives a retrieval system a reason to trust neither.

Does the description contain checkable facts, or only adjectives? This is the difference that decides citation more than any technical setting. Exceptional service in an idyllic setting says nothing a system can verify or reuse. Twelve rooms, open year-round, fifteen minutes from the airport at La Tontouta, English and French spoken, says four things that can be checked and quoted. Promotional language without supporting fact is exactly what a careful editorial process strips out.

Are place names spelled correctly, with accents intact? Noumea, Dumbea, Mont-Dore, Paita, Lifou, Mare, Ouvea, the Isle of Pines, Grande Terre. These are identifying tokens. A misspelled or de-accented place name weakens the association between your record and the place it describes, and in a territory where the local content pool is thin there is no volume of other signals to compensate.

Are Kanak and Melanesian terms preserved rather than normalised? The territory has 28 Kanak languages, per the Vice-Rectorat language service, with approximately 76,000 speakers aged 15 and over recorded at the 2019 census, plus a further 17,000 who understand one without speaking it. Local terminology, including words like coutume and tribu that carry specific meaning here, should survive editing rather than being smoothed toward metropolitan French. Careful local editorial practice preserves these deliberately, and the practice is right for reasons beyond accuracy: those terms are part of how a system recognises that a page is about this territory rather than about France.

Do your record and your own site agree? Disagreement between two sources about the same business is one of the surer ways to be left out of an answer. If Apidae says you close on Sunday and your site says you open, a system has no basis for choosing and reasonable grounds for citing something else.

Does your description read like everybody else's? Since Apidae introduced generative description tools in October 2024, the risk of homogenisation has risen. A record that could describe any of two hundred similar businesses gives a system no reason to select yours specifically.

What does not help, despite what you have been told

Two claims circulate widely enough that leaving them unaddressed would be a disservice, and both are easier to disprove than to argue about.

The first is that special schema markup is required to appear in AI features. Google states plainly that no special markup is needed for its generative AI surfaces. Implement structured data anyway, because it genuinely helps conventional search and machine parsing, but do not buy it as an AI unlock and do not sell it as one. The distinction matters commercially: a vendor charging a premium for AI schema is charging for a mechanism the platform says does not exist.

The second is llms.txt, a proposed file for declaring content to language models. It is not a documented requirement of any major platform. What is documented is crawler access, and that is the check actually worth running: whether the retrieval agents you care about, including OAI-SearchBot for inclusion in ChatGPT search, are permitted to reach your pages at all. A robots directive quietly blocking a retrieval agent does real damage. A missing llms.txt does not.

There is a useful contrast here with a source that does carry weight. The French government's digital resource portal for businesses publishes guidance on generative engine optimization aimed at tourism professionals, treating establishment records, practical information, opening hours and descriptive content as strategic elements for appearing in AI-generated answers. That is a public administration, not an agency selling a service, arriving at the same conclusion about records rather than markup.

Claim	Status	What to do instead
Special schema markup unlocks AI features	Contradicted by Google's own guidance	Implement schema for conventional search and parsing, without the AI promise attached
You need an llms.txt file	Not a documented requirement anywhere	Verify crawler access, including OAI-SearchBot, is not blocked
More links from anywhere improves citation	Weak here specifically	Two local publications carry real weight; volume placement is visible and ineffective in a pool this small
Translate the French national page and localise later	Loses to locally written content	Write for the territory, preserving Caledonian place names and Kanak terms
A generic industry record is enough to be found	Fails on ambiguity	Include checkable facts, correct place names and details that could only describe your business

Where the vocabulary matters

One more layer sits underneath all of this, and it is the one an outside agency is most likely to flatten by accident.

Caledonian French carries its own vocabulary. La metropole means mainland France. Le Caillou, literally the pebble, is what residents call the territory itself. La Brousse refers to everything outside greater Noumea. Coutume and tribu carry specific local meaning that has no clean metropolitan equivalent. These are not regional quirks to be smoothed out in editing; they are how the place talks about itself.

For retrieval, that has a direct consequence. A page using this vocabulary correctly reads as being from and about this territory. A page that has replaced it all with standard metropolitan phrasing reads as being about France with the place name changed, which is precisely the substitution the generic prompt produced in our probe.

Get it wrong in the other direction too and the damage is worse. Vocabulary used incorrectly by someone who does not know the territory is more visible to a local reader than absence would have been, and it undermines the authority the content was written to build.

Tessar Napitupulu writes about entity clarity and structured sources as the foundation of citation in Cited or Silent, available as a free gated edition, with retailer editions on Amazon, Google Play and Apple Books. Record hygiene of the kind described here sits inside scope for tourism and hospitality clients of our GEO service for New Caledonia, rather than being treated as somebody else's paperwork.

---

Frequently Asked Questions

What is Apidae and why does it matter for AI visibility?

Apidae is the shared structured tourism database used across France and its overseas territories, with more than 25,800 users across the network. New Caledonia has run on it since 2019, with Sud Tourisme Nouvelle-Caledonie maintaining roughly 3,200 destination records on a daily basis through the tourist office adviser network. Those records feed the destination website, events listings, print and digital promotion, and partner sites. It matters because a consistent, machine-readable, institutionally maintained source is exactly the kind of thing answer engines favour, and this territory already funds and staffs one.

Why do generic French prompts return metropolitan results for this territory?

Because a retrieval system does not return nothing when it lacks local sources, it returns the nearest available thing, and for a French-speaking territory that is France. In a limited probe in August 2026, a generic prompt about the best GEO agency in New Caledonia surfaced metropolitan hosting and marketing pages plus local businesses matched only on the word agence, including a real estate agency and a surveying practice. A more specific prompt about AI visibility for a company in New Caledonia retrieved a genuinely local page. That is suggestive of a conditional pattern rather than proof of one, and three prompts cannot establish how AI systems treat a whole territory.

What should I check on my Apidae record?

Six things. Whether the record exists and is current today rather than in 2023, since a great deal changed after May 2024. Whether the description contains checkable facts rather than only adjectives, because room counts, opening periods and distances can be verified and reused while exceptional service cannot. Whether place names are spelled correctly with accents intact. Whether Kanak and Melanesian terms survive editing rather than being normalised. Whether the record and your own website agree, since disagreement gives a system grounds to cite neither. And whether the description could describe any of two hundred similar businesses, which since generative description tools arrived in October 2024 has become a real risk.

Do I need special schema markup to appear in AI answers?

No. Google states that no special markup is required for its generative AI features. Implement structured data anyway, because it genuinely helps conventional search and machine parsing, but treat any vendor charging a premium for AI schema as charging for a mechanism the platform itself says does not exist. The same applies to llms.txt, which is not a documented requirement of any major platform. What is documented and worth checking is crawler access, including whether agents such as OAI-SearchBot are permitted to reach your pages.

Why do Kanak place names and terms matter technically?

Because they are identifying tokens. The territory has 28 Kanak languages, per the Vice-Rectorat language service, with approximately 76,000 speakers aged 15 and over recorded at the 2019 census, plus a further 17,000 who understand one without speaking it, and local terminology including words like coutume and tribu carries specific meaning with no clean metropolitan equivalent. A page that preserves this vocabulary reads as being from and about this territory. A page that has replaced it with standard metropolitan phrasing reads as being about France with a place name substituted, which is the exact confusion that produces metropolitan results for local queries. Careful local editorial practice preserves these deliberately.

Is link building worth doing in a market this small?

Yes, but not by volume, and the arithmetic is unusual. The pool of genuinely local publishers that carry weight is essentially two, Les Nouvelles Caledoniennes and NC la 1ere. That cuts both ways: each placement counts for considerably more than it would in a large market, and there is no volume strategy available to hide behind. Institutional sources matter as much, since correctly citing ISEE, IEOM, the CEROM partnership and gouv.nc connects your content to publishers a retrieval system already treats as authoritative for this territory.

---