Availability Is Not Frequency In AI Search

Google's own coverage documentation lists Tonga among the countries where AI Overviews are available. That is a documented fact from a primary source, and it is the single most commonly misused fact in proposals written for small markets.

Because here is what it does not say. It does not say how often an AI Overview appears when someone searches something about Tonga. It does not say what gets cited when one does appear. It does not say whether the feature behaves the same way for a query typed in Nuku'alofa as for the same query typed in Auckland. Every one of those is a separate question, and none of them has a published answer.

Availability is a fact about a feature. Frequency is a fact about a market. Converting the first into the second is the most common piece of dishonesty in this industry, and it usually happens by implication rather than by outright claim.

How the conversion happens

Nobody writes "AI Overviews trigger 40% of the time in Tonga." That would be checkable and it would be wrong.

What happens instead is a sequence. The proposal states that AI Overviews are available in Tonga, which is true. It then cites a global trigger rate from a study conducted somewhere else, which is also true in its own context. It places the two adjacent, in consecutive sentences, and lets the reader do the arithmetic. Nothing false has been written. A false impression has been created.

Watch for the same pattern with adoption figures. During research for this project we found three separate statistics about AI use in travel planning, all real, all from named sources, and all measuring different things: roughly 30% of United States travellers using AI extensively for trip planning in 2025, up from 13% the previous year; nearly 40% of United States travellers using generative AI for travel research; and 56% of United States travellers using AI to plan, book or accompany at least one trip. Those are three different behaviours with three different thresholds. Presented as a range, they suggest a measurement. They are not a measurement, they are three studies asking three questions, and three of the four research passes for this project flagged the merging risk independently.

None of them measured Tonga. All of them measured the United States.

The one strong sector statistic, and its three conditions

There is a genuinely useful figure here, and it deserves to be used properly rather than stripped for parts.

Similarweb's 2026 generative AI benchmark report found that travel carries the highest citation share of the nine sectors it tracked: 22.6% of ChatGPT answers in the travel category contained at least one citation, against a cross-sector average of 6.8%. Retail came second at 13.5%, then sports at 10.7%, finance at 8.0%, technology at 6.6%, health at 6.3%, media and entertainment at 6.0% and education at 4.8%. The overall average had risen from somewhere around 1.3% to 1.6% in June 2025.

Three conditions travel with that number, always.

The geography is the United States, on desktop, for May 2026. Not global. Not New Zealand. Certainly not Tonga.

These are modelled estimates, not measured platform data. Similarweb's own disclaimer describes its data as estimates and extrapolations from third-party sources, without warranty of accuracy. Similarweb also sells AI visibility tracking products. That does not make the figure wrong, and the report is one of the better public datasets available. It does mean the correct confidence marker is reported rather than verified.

It measures whether any citation appeared, not who won it. This is the condition most often dropped and it is the most consequential. A 22.6% citation share in travel tells you the citation layer in that category is comparatively wide. It tells you nothing whatsoever about whether a small operator can get into it. Those are different claims and only the first one has evidence behind it.

Used with all three conditions attached, the figure supports a real argument: travel is a category where assistants cite sources more often than they do elsewhere, so publishing citable material has a better prior in travel than in most industries. That is a reasonable basis for doing the work. It is not a forecast.

Why your KPI might be measuring someone else's release notes

Here is the finding that should decide how a Tonga engagement is reported.

Following a ChatGPT update on 7 May 2026, the share of AI referral traffic landing on homepages rather than deeper pages rose from roughly a quarter to nearly sixty percent within about three weeks. No brand did anything. No content changed. A platform shipped an update and a widely tracked metric moved by more than thirty points.

Sit with what that means for reporting. If your success measure is built on referral behaviour, then a substantial share of your reported movement is platform behaviour, and you cannot separate the two after the fact. A client looking at a chart showing homepage referrals doubling has no way to know whether that was your work or a release on 7 May.

The same report noted that around 65% of cited URLs sat two or three folders deep, while 58.8% of AI referral traffic landed on homepages. Cited pages and landing pages are not the same set. Optimising the second while reporting it as evidence about the first is a category error, and it is easy to make accidentally.

The response is not to give up on measurement. It is to measure something that platform updates cannot silently rewrite.

A fixed prompt panel, and why we state the denominator

The method is unglamorous and it works.

Write a fixed set of prompts that a real customer would actually type, in the language and phrasing they would use. Run them across the assistants that matter, on a stated schedule, from a stated location. Record whether the business appears in the citation set, as a count out of the panel size. Re-run the identical panel at the same interval. Report the change in counts.

That last point about the denominator matters more than it sounds. There is no industry standard setting a minimum panel size for this kind of reporting, and we checked. One research pass proposed a fixed benchmark bank of a hundred prompts and presented it as a standard, which it is not, and no source could corroborate it. Another suggested twenty to thirty. A third stated plainly that no authoritative standard exists.

So we state our panel size rather than implying a norm. Twelve out of forty prompts is a reportable finding. Thirty percent, without the forty, is a number that could mean anything, and on a small market it usually means very little.

Metric	Stable against platform updates?	Use it as
Citations out of a fixed prompt panel, stated size	Largely yes, if the panel and method are unchanged	Primary measure
Per-platform citation counts, reported separately	Yes, and it exposes divergence a blended score would hide	Primary measure
Factual accuracy of what the assistant says about you	Yes, and it is often the fastest thing to improve	Primary measure
Enquiries, booking requests, booking value	Yes. It is your own ledger	Outcome measure
AI referral traffic volume	No. Moved by more than thirty points in three weeks on a single platform update	Context only, never a target
Blended cross-platform visibility score	No, and it conceals the gap you are paying to find	Avoid
Percentage change without a denominator	Meaningless at low volumes regardless of platform behaviour	Never

The percentage problem, stated once and properly

Two citations becoming four is a 100% increase. It is also two citations.

On a market of 104,000 people with roughly 60,700 internet users, low absolute numbers are the normal condition rather than a sign of failure. Percentage reporting on those bases produces charts that look dramatic and mean almost nothing, and clients work this out eventually. When they do, everything else in the report becomes suspect too.

Every independent research pass conducted for this project reached this conclusion separately, across both the search and the answer-engine tracks. That kind of unanimity is unusual enough to be worth acting on. Report counts. Where a percentage genuinely helps, publish the numerator, the denominator and the measurement window beside it, every single time.

What a first engagement honestly looks like

Measurement first, recommendations second. In that order, because reversing it means guessing.

The opening deliverable is a baseline: a fixed prompt panel, run across the assistants, from the origin markets that matter, with the result recorded as counts and the accuracy of each mention noted. That produces a document showing where the business currently sits, which nobody has ever produced for it before, because no such baseline exists for any Tongan business.

What comes next depends on what the baseline shows, and the honest possibilities include outcomes that reduce the scope of work. If the business is already accurately represented in most of the panel, the remaining work is maintenance rather than a programme. If it is absent everywhere and the reason is that the assistants have no source to draw on, the answer is publication rather than optimisation. If it appears but with wrong details, the fix is entity correction, which is fast and cheap.

None of those conclusions can be reached in advance. That is the point of measuring first, and it is why we will not write a strategy document for a Tongan business before we have run the panel.

Tessar Napitupulu sets out the measurement discipline behind answer-engine work, including why absolute counts beat rates on small bases, in Cited or Silent.

---

Frequently Asked Questions

Are AI Overviews available in Tonga?

Yes. Google's own coverage documentation lists Tonga among the countries where the feature is available, which is a documented fact from a primary source. That is a statement about availability only. How often it triggers for Tonga queries, and what it cites when it does, is absent from every source reviewed for this article.

What is the difference between availability and frequency?

Availability is a fact about a feature being switched on in a country. Frequency is a fact about how often it actually appears for real queries in that market. Proposals commonly state the first, cite a global trigger rate measured somewhere else in the next sentence, and let the reader infer the second. Nothing false is written and a false impression is created.

Is the 22.6% travel citation figure reliable?

It is one of the better public datasets available, with three conditions that must travel with it. The geography is United States desktop data for May 2026, not global and not Tonga. The figures are modelled estimates extrapolated from third-party sources, published by a vendor that also sells AI visibility tracking, so the correct marker is reported rather than verified. And it measures whether any citation appeared in a category, not whether a particular business can win one.

Why not report AI referral traffic?

Because it moves on platform updates you do not control. Following a ChatGPT update on 7 May 2026, the share of AI referral traffic landing on homepages rather than deeper pages rose from roughly a quarter to nearly sixty percent within about three weeks, with no action taken by any brand. Reporting built on referral behaviour cannot separate your work from a release note after the fact.

How many prompts should a tracking panel contain?

No authoritative standard exists, and we checked. One research source proposed a fixed bank of a hundred prompts and presented it as a standard, which no source could corroborate. Another suggested twenty to thirty. A third confirmed plainly that no industry minimum has been established. The workable answer is to state your panel size in every report rather than implying a norm, so that twelve out of forty is reportable and thirty percent on its own is not.

Why report counts instead of percentages?

Because two citations becoming four is a 100% increase and also two citations. On a market with roughly 60,700 internet users, low absolute numbers are the normal condition rather than a sign of failure, and percentage reporting on those bases produces charts that look dramatic and mean very little. Every independent research pass conducted for this project reached the same conclusion separately across both tracks.

Should visibility across platforms be combined into one score?

No. A business can appear in one assistant and be entirely absent from another on the identical prompt, and a blended score conceals exactly the gap the engagement is meant to find. Report each platform separately, with its own count and its own denominator.

What does a first engagement produce?

A baseline before any recommendations: a fixed prompt panel run across the assistants from the relevant origin markets, with results recorded as counts and the factual accuracy of each mention noted. No such baseline currently exists for any Tongan business. What follows depends on what it shows, and the honest possibilities include findings that reduce the scope of work, such as the business already being accurately represented and needing maintenance rather than a programme.