Are AI Overviews Live In New Caledonia

For sixteen months, French speakers in New Caledonia could reach a Google feature that French speakers in Paris could not. That sentence sounds wrong, and it is the sort of thing that gets corrected in a meeting by someone confident that an overseas collectivity of the French Republic must sit inside whatever perimeter France sits inside.

It does not. Google maintains one availability list for AI Overviews, and New Caledonia appears on it in the Asia-Pacific column, alongside Australia, New Zealand, Fiji and Vanuatu. France appears in a different column. Access is determined by a user's physical location, not by the country attached to their Google account. So the position of this territory on Google's AI map is set by its geography, not by its constitutional status.

That is the first of two gates. The second is language, and it opened on a completely different schedule. Understanding that both exist, and that they moved independently, is the difference between planning for this territory accurately and assuming it behaves like a French department.

Gate one: is the territory on the list

Territory availability is a published fact rather than an inference. Google's support documentation lists the countries and territories where AI Overviews are available to all users, and New Caledonia is on it.

What makes the placement interesting is the neighbourhood. In Google's grouping, this territory sits with the Pacific rim rather than with Europe. French Polynesia, Wallis and Futuna, Vanuatu, Fiji, Papua New Guinea, Australia and New Zealand are the peers. Metropolitan France is in the Europe, Middle East and Africa group.

Combine that with the location rule and the consequence is concrete. A user searching in Noumea and a user searching in Paris are, for this purpose, in different buckets that were switched on at different times. Their experience of the same query in the same language could differ for reasons that have nothing to do with the query or the language.

Anyone planning here should check the current list rather than trusting an article, including this one. Availability lists change, and the honest instruction is to verify rather than to cite.

Gate two: is the language supported

This is the gate most people do not know exists, and it explains the sentence this article opened with.

When AI Overviews expanded to more than a hundred countries in October 2024, the supported languages were English, Hindi, Indonesian, Japanese, Portuguese and Spanish. French was explicitly not among them. So a territory could be inside the availability perimeter while its working language was outside it, and users would experience that as the feature not really being there.

French arrived later, on its own schedule. In the AI Mode expansion of October 2025, which took that feature past 200 countries and territories, French was added alongside Italian, German, Dutch, Polish and Greek. Note what was happening simultaneously: France itself was still excluded from both features while its language was being switched on for everyone else.

Two gates, two schedules, and no requirement that they open in any particular order. That is the structural insight, and it generalises beyond this territory to every market whose language and geography do not line up neatly.

Why France waited, and what it cost

France was the last major Western European market without AI Overviews or AI Mode. The reason was regulatory rather than technical, and it is worth understanding because it shaped the timeline for every French-language market including this one.

The dispute concerned neighbouring rights, the mechanism by which French publishers are compensated when their content is used. Google had already been fined 500 million euros in 2021 over its conduct in those negotiations, and a further 250 million euros in 2024. Against that backdrop, launching a feature that summarises publisher content into an answer box was, from Google's position, a legal exposure rather than a product decision.

Google's senior leadership said as much publicly. In October 2025, while announcing the AI Mode expansion that added French as a language, the company stated openly that it hoped to resolve the regulatory uncertainty in France that had made features like AI Overviews and AI Mode much harder to launch there, and that it was not yet clear when that would be possible. In May 2026, Google Product Experts were still confirming in support threads that the feature was unavailable in France due to regional restrictions.

The resolution came on 22 July 2026. Google switched on both AI Overviews and AI Mode for French searchers across desktop, mobile web and the Google app, announced on its French blog with the VP of Search making the announcement publicly the same day. Two days before that, on 20 July 2026, Google had published Search Console documentation for a control letting site owners exit AI Overviews and AI Mode, formalising a mechanism it had begun testing in June.

The rollout is progressive rather than instant, with full query coverage targeted toward the end of September 2026. That matters for measurement, and we come back to it.

What the first measured French impact actually shows

Until August 2026 there were no French numbers, only borrowed ones. On 17 August 2026 that changed, with an analysis of 963 domains drawn from the top thousand click-generating domains in France, using aggregated Google Search Console data across 28 days before the launch compared with 9 days after it.

The split by exposure is the useful part.

Group	Definition	Click-through rate change	Share of group losing CTR
High exposure	More than 20% of queries triggering an AI Overview	Down 23.1% on average	82.2%
Control	Under 2% of queries triggering an AI Overview	Up 2.3%	Not applicable

Three caveats travel with that table, and dropping any of them turns a useful measurement into a misleading one.

Nine days is a very short window. It was measured mid-rollout, when coverage was still expanding toward the end-September target, so the exposure classification was moving underneath the measurement. And it is France, not New Caledonia. No equivalent analysis has been published for this territory, and given the data-coverage problems that affect this market generally, none is likely soon.

For a longer-run comparison, Germany offers one. Documented in March 2026, roughly 265 million monthly organic clicks were estimated lost, with click-through rate at the first organic position falling from 27% to 11%. Germany received AI Overviews in March 2025, so that is a full year of exposure rather than nine days. It is also a different market with different publisher dynamics, and it should be read as an upper-bound illustration rather than a forecast.

The defensible summary for a Caledonian business is short. The effect is real, it concentrates on sites whose queries trigger AI answers frequently, and the magnitude for this territory is unmeasured. Plan for direction, not for a number.

What this changes about measurement here

A moving feature makes trend lines unreliable, and that is the practical trap in this period.

If you start monitoring in August 2026 and compare against November 2026, you are measuring two things at once: whatever your own work achieved, and however much further the rollout progressed in between. Attribution between those two is not possible after the fact. So the discipline is to date every measurement, hold the prompt set fixed, and repeat on a schedule rather than opportunistically.

Three specific habits follow.

Record the raw output, not just a score. When a model updates and your score moves, the raw citations tell you whether the sources changed or only the phrasing did. A score alone cannot distinguish those.

Run the same prompts, in the same language, with the same wording. Small rewordings produce different retrieval, which is a finding in itself but ruins a time series. Keep a separate exploratory set if you want to test variations.

State the rollout stage in every report. A citation share measured in August 2026 and one measured in October 2026 sit on different sides of the coverage target, and a reader six months later needs to know that without having to reconstruct it.

Google is one surface, and not the largest behaviour change

Everything above concerns one feature from one company. Framing the whole question around AI Overviews understates part of the shift and overstates another part, and the French national data lets us be precise about both.

The Barometre du numerique, run by CREDOC for Arcep, Arcom, the CGE and the ANCT, surveyed 4,145 people aged 12 and over and published its 2026 edition in February. It found that 48% of French people use generative AI, up from 20% in 2023. In twenty-five years of the survey, no digital technology had been adopted that fast. About a third of users reach for it daily, rising to 51% among 18 to 24 year olds, a group where overall adoption runs at 85%.

Usage concentrates heavily. ChatGPT is the main service for 63% of French AI users, ahead of Gemini at 13% and Le Chat, the French-built assistant, at 6%. For a business deciding where to start monitoring, that concentration is good news: a narrow prompt set on one platform still covers most of the exposure.

Now the corrective, and it is the figure most likely to be left out of a vendor pitch. For finding information specifically, French users still turn to search engines roughly twice as often as to generative AI, 59% against 28%. Generative AI has overtaken other tools for writing and translation at 49%, content creation at 49% and programming help at 48%, but not for the behaviour that search marketing depends on.

Read those two facts together and the honest position falls out. Adoption is fast and concentrated, so ignoring the answer layer is a mistake. Search is still the dominant information-seeking behaviour by a factor of two, so moving budget wholesale out of search into AI visibility is a bigger mistake. Build both, with the answer layer designed into the ranking work rather than replacing it.

One more finding is worth carrying into content decisions. The most cited reason French people give for not using generative AI is lack of trust, either in how their data is handled or in the answers themselves, at 30%. Habit accounts for 28% and lack of skill for 26%. If distrust of answers is the largest barrier, then sourced, checkable, non-promotional writing is not only what earns a citation, it is also what earns the reader who arrives through one.

The obvious caveat applies to all of it. These are French national figures from a French national survey. There is no equivalent study for New Caledonia, and nobody should present these percentages as Caledonian. They are the nearest defensible proxy, used as a proxy and labelled as one.

What a Caledonian business should do about it

Four things, none of which require knowing the unmeasured territory number.

Measure your own exposure before anything else. The French analysis found the entire effect concentrated in one group: sites where AI answers appeared on more than a fifth of queries. Sites with almost no exposure gained slightly. You can determine which group you are in from your own data, and that single fact should drive the size of your response. A business whose queries rarely trigger an answer box has a different problem from one whose queries almost always do.

Do not treat the opt-out as a strategy. The Search Console control documented on 20 July 2026 lets a site owner exit AI Overviews and AI Mode. It exists, it is legitimate, and using it removes you from a surface where your competitors will still appear. In a territory with a thin pool of local content, being absent is a stronger signal than being summarised. Consider it a tool for specific pages with specific problems, not a posture.

Write for citation, not for the summary you fear. The mechanism that gets a page cited is the same one that makes it useful to a reader who arrives directly: a clear answer near the top, facts that are attributable and checkable, headings that match the question as asked. That work pays whether or not an answer box appears above it.

Keep the French question in front of you. Both gates are open here now, and one of them is a language gate. Content in weak French is content that a retrieval system has grounds to pass over in favour of a metropolitan source written by a native speaker. In a market where the pool of local French content is small, there is not enough volume for that disadvantage to average out.

Tessar Napitupulu writes about the difference between ranking and being cited, and how to measure the second one, in Cited or Silent, available as a free gated edition, with retailer editions on Amazon, Google Play and Apple Books. The dated, fixed-prompt measurement discipline described here is how our GEO service for New Caledonia reports during a rollout that has not finished.

---

Frequently Asked Questions

Are Google AI Overviews available in New Caledonia?

Yes. Google's published availability documentation lists the territory, and French is a supported language. The detail worth knowing is the grouping: New Caledonia appears in the Asia-Pacific column alongside Australia, New Zealand, Fiji, Vanuatu, French Polynesia and Wallis and Futuna, not in the European column with France. Access follows a user's physical location rather than the country attached to their Google account. Availability lists change, so verify against the current documentation rather than relying on any article, including this one.

How could a French territory have the feature before France did?

Because territory availability and language availability are two separate gates on two separate schedules. New Caledonia sits in Google's Asia-Pacific grouping, which was inside the availability perimeter from the wider expansions of October 2024 and May 2025. France was explicitly excluded from the European expansion of 26 March 2025 and stayed outside until 22 July 2026. Meanwhile French was absent from the October 2024 launch languages and only became supported in the October 2025 expansion. Neither gate waits for the other.

Why was France excluded for so long?

A neighbouring-rights dispute with French publishers, not a technical constraint. Google had already been fined 500 million euros in 2021 and a further 250 million euros in 2024 in related proceedings. In October 2025 the company said publicly that it hoped to resolve the regulatory uncertainty in France that had made these features harder to launch, and that it was not yet clear when that would be possible. Google Product Experts were still confirming unavailability in May 2026. The launch followed on 22 July 2026.

How much traffic does an AI Overview actually cost a site?

In France, the first measured answer came on 17 August 2026 from an analysis of 963 top click-generating domains using Search Console data across 28 days before the launch against 9 days after. Sites where more than 20% of queries triggered an AI Overview lost 23.1% of click-through rate on average, with 82.2% of that group losing CTR. A control group with under 2% exposure gained 2.3%. Nine days is short, it was measured mid-rollout, and it is France rather than New Caledonia.

Is there any New Caledonia specific measurement?

No. No study has published AI answer prevalence, citation share or click-through impact for this territory, and given the wider data-coverage problems in this market none is likely soon. The French analysis is the nearest evidence and it describes France. A longer-run comparison exists for Germany, where roughly 265 million monthly organic clicks were estimated lost and first-position click-through fell from 27% to 11% after a full year of exposure, but that is a different market and should be read as an upper-bound illustration rather than a forecast.

Should we use the Search Console opt-out?

Rarely, and never as a general posture. Google documented the control on 20 July 2026, formalising a mechanism tested from June, and it does let a site owner exit AI Overviews and AI Mode. But opting out removes you from a surface where competitors will still appear, and in a territory with a thin pool of local content, absence is a stronger negative signal than summarisation. Treat it as a page-level tool for specific problems rather than a site-wide decision.

How should we report results while the rollout is still moving?

Date everything, fix the prompt set, and repeat on a schedule. Coverage was still expanding toward an end-September 2026 target, so a comparison between August and November measures both your own work and the rollout's progress with no way to separate them afterwards. Record raw citation output rather than only a score, because when a score moves the raw output tells you whether the sources changed or only the phrasing. And state the rollout stage in every report, so a reader six months later does not have to reconstruct it.

Is search being replaced by AI assistants in French-speaking markets?

Not yet, and the gap is larger than most vendor pitches suggest. The 2026 Barometre du numerique, run by CREDOC for Arcep, Arcom, the CGE and the ANCT across 4,145 people aged 12 and over, found 48% of French people using generative AI, up from 20% in 2023 and the fastest adoption the survey has recorded in twenty-five years. But for information seeking specifically, search engines are still used roughly twice as often as generative AI, at 59% against 28%. Generative AI has overtaken other tools for writing and translation, content creation and programming help, all near 49%, which are different behaviours from search. Build both layers; do not move budget wholesale from one to the other.

---