Reporting When Every Number Is in the Tens

A tourism operator on Saint Helena checks the analytics. Organic sessions went from eight last month to twelve this month. The dashboard renders that as a fifty percent increase and colours it green. Somewhere a slide gets made.

What actually happened, in all likelihood, is that one family in Hampshire planned a trip using a laptop, a phone and a tablet, and one of them came back twice. Nothing improved. Nothing got worse. The reporting simply described household browsing behaviour in the language of growth.

This is not unique to Saint Helena. It is a small-numbers problem, and Saint Helena is the most extreme small-numbers market most agencies will ever meet: 3,980 residents at end June 2025, 889 non-St Helenian leisure visitors in the twelve months to October 2024, nominal GDP of £40.7 million for the whole territory. At that scale most reporting conventions stop working, and some do worse than stop. They mislead.

The useful news is that a workable method exists, and the territory's own government already uses it. What follows is that method, why the alternatives fail, and where measurement genuinely runs out.

Four arrival numbers, none interchangeable

Before anything can be measured, the denominator has to be settled, and Saint Helena has four plausible ones. The Statistics Office is careful about this. Almost nobody downstream is.

The tourist figure is 889. That means non-St Helenian arrivals whose stated purpose was leisure, over the twelve months to October 2024, up from 857 in the comparable prior window. Widen the definition to all leisure arrivals, including Saint Helenians returning for a holiday, and the same period gives 1,983. Widen it again to every arrival of any purpose, and the twelve months to June 2025 produced 5,298, against 4,344 in the preceding period. Narrow instead to a single month and October 2024 shows 149 leisure arrivals.

Every one of those is accurate. Presented without its definition, any of them can be passed off as the visitor number, and the gap between the smallest and the largest is a factor of nearly six. A campaign judged against 5,298 looks feeble. The same campaign judged against 889 looks transformative. Neither judgement means anything until somebody says which figure they used and why.

One further wrinkle. Because releases use different twelve-month windows, two correct figures from the same office can look contradictory purely because one ends in October and the other in June. Fix the window and hold it.

Figure	What it counts	When to use it
889	Non-St Helenian leisure visitors, twelve months to October 2024	Tourism marketing performance. This is the addressable market for destination work
1,983	All leisure arrivals including returning Saint Helenians, same window	Accommodation and hospitality demand, where a returning resident still books a room
5,298	Total arrivals of every purpose, twelve months to June 2025	Airport and logistics capacity planning. Almost never the right marketing denominator
149	Leisure arrivals in October 2024 alone	Illustrating seasonality. Never for a trend, because monthly variance swamps any signal
2,481	Tourists in 2019, the pre-pandemic peak	Recovery framing. Useful context, and a reminder of how far the baseline has moved
3,980	Island residents at end June 2025	Anything aimed at residents. Not to be mixed with combined-territory figures

Pick one, write it at the top of the report, and never quietly swap it. Half the disputes about performance in markets like this are really disputes about denominators that nobody named out loud.

Why the standard dashboard is worse than nothing here

A conventional monthly report contains rank positions, session counts with period-on-period percentages, conversion rates, and increasingly some measure of AI visibility expressed as a share. Take each in turn.

Rank positions assume a stable pool of competing pages and enough query volume for the ranking to be sampled meaningfully. Neither holds. Keyword tools round small numbers toward zero here, so the rank being tracked belongs to a query nobody can confirm anybody searched. Session percentages have already been dealt with: on a base of tens they measure household device counts.

Conversion rate is the quiet disaster. With three enquiries in a month, 4.2 percent is not a rate, it is one division performed on a sample too small to have one. Worse, it invites optimisation, and at that volume the only way to move it is luck.

Share of voice in AI answers is the newest and the most seductive. It requires a denominator, and no published source provides one for this territory. Any percentage of AI visibility quoted for Saint Helena has a denominator somebody chose. Ask what it was and the conversation usually ends.

None of this means measurement is impossible. It means the unit is wrong. Reporting defaults to rates because rates are comparable across accounts, and comparability is exactly what has to be sacrificed here.

Borrow the method that already works on the island

The St Helena Statistics Office publishes regular updates on arrivals and departures, and its conventions are a ready-made template. It reports absolute counts. It compares twelve-month rolling periods rather than months. It states the definition attached to each figure. It marks provisional data as provisional. It has been doing this for years, against a visitor base of a few hundred, without once needing to report a percentage swing on a monthly base.

A tourism arrivals tracking system, T-Stats from Acorn Tourism, has been in place since 2017. The client is already fluent in this style of reporting, so handing them a marketing report built on different conventions creates friction for no gain. Adopting the same conventions makes marketing performance directly comparable to arrivals data, removes the temptation to headline a movement four people caused, and aligns the reporting period with an annual rather than monthly planning cycle.

What to count instead

Substitute counted events for calculated rates, and pick events that a single household cannot manufacture.

Enquiry and booking-start counts, absolute, twelve-month rolling. Not conversion rate. Not month on month. The number of enquiry forms, the number of booking journeys started, the number of brochure downloads, each as a plain integer with a rolling twelve-month total beside it.

Coverage counts. Number of distinct third-party pages that describe the territory accurately, number that describe it inaccurately, number of correction requests sent and accepted. These are countable, attributable to work, and unaffected by traffic volume.

Error register size. How many factual errors about access, permits, currency or contacts currently exist across the retrievable corpus, and how that count changes. This is the metric that behaves best in this market, because it is a direct function of work done rather than of anybody's browsing.

Indexation and freshness counts. Number of key pages with a visible effective date. Number of superseded documents still retrievable. Number of contradictions found and resolved within the owned corpus.

Tender and procurement events. For public-sector work, registrations on the procurement portal, expressions of interest, invitations received. Small integers, high information content.

Citation observations. For AI visibility, the count of runs in which the territory was described accurately, out of a fixed number of runs. Sixteen out of twenty is a meaningful sentence. Eighty percent citation rate, from a sample of twenty, invites a precision the sample cannot support.

The repeat-run problem, and why a screenshot proves nothing

AI visibility reporting has a specific instability that small markets feel hardest.

In January 2026 SparkToro, working with Gumshoe.ai, had roughly 600 volunteers submit twelve prompts across ChatGPT, Claude and Google AI systems, producing 2,961 recorded runs. Search Engine Land's coverage put the odds of receiving the same list of recommendations twice at under one in a hundred, and the same list in the same order at closer to one in a thousand.

Read that again, because it has an uncomfortable implication for a common practice. A vendor showing you a screenshot in which your brand appears in an AI answer has shown you one draw from a distribution with roughly a one percent chance of repeating. It is evidence of possibility, not of position. The same is true in reverse: an absence in one run is not an absence.

The only defensible response is repetition under controlled conditions. Fixed prompt wording, the same declared location settings, the same engines and modes, run enough times that the distribution becomes visible rather than one sample of it. Then report accurate runs as a count out of the total, and log the sources named each time.

A 2026 critical survey of the generative-engine-optimisation literature, reviewing forty-five studies, made a related point that deserves more attention than it gets: discoverability, retrieval, citation, prominence, referral traffic and commercial outcome are separate variables, and the published evidence does not establish that improving one reliably improves the next. Collapsing them into a single visibility score is convenient and unsupported.

There is also a structural finding worth knowing before setting expectations. A 2026 study published in Current Issues in Tourism examined 420 travel recommendations generated by ten AI systems across fourteen structured queries, and found that while the surface variety looked broad, attention remained concentrated on destinations that were already highly visible. For a territory with 889 leisure visitors a year, that is the headwind. It does not make the work pointless. It does mean the realistic goal is accuracy when asked about, rather than unprompted recommendation.

A cautionary tale in third-party numbers

One more hazard, because it undermines reports from the outside rather than the inside.

DataReportal publishes annual digital adoption figures covering this territory. Across three editions, internet users appear as 2,913 and 55.2 percent in early 2024, then 3,300 and 63.3 percent in January 2025, then 1,950 and 37.6 percent in October 2025. Social media users move similarly. Those swings exceed forty percentage points in under two years, in a place where the resident population changed by a few dozen people.

Population did not move. Methodology did. The figures also cover the combined territory rather than the island alone, which is a different geography from the one the Statistics Office reports on. Meanwhile the 2021 census, measuring households rather than individuals on the island only, found 66 percent with internet access at home.

Three lessons. Never mix a household measure with an individual one. Never mix island-only with combined-territory scope. And never build a trend line across editions of a third-party estimate whose method changed, because the chart describes the publisher rather than the place.

Write it into the contract

Measurement conventions belong in the agreement, not in a footnote to month three's report. Four clauses do most of the work.

Name the denominator and the window. Which arrival figure, which twelve months, and a commitment not to change either without a written note explaining why.

Define what counts as an observation for AI reporting. Number of runs, engines, modes, declared location, prompt wording frozen for the term. If the panel changes, the old panel keeps running alongside for one cycle.

State explicitly what will not be reported, and why. This reads as unusual and it prevents the slow drift toward vanity metrics that happens when nobody wrote down what was excluded.

Set a review point where continuation is genuinely in question. An engagement that cannot say what would justify stopping is not being measured. It is being renewed.

Where measurement genuinely runs out

Some things cannot be measured here, and saying so is part of the method rather than a failure of it.

Attribution from an AI answer to a confirmed booking cannot be done at this volume. Referral data from AI interfaces is sparse and inconsistent even in large markets, and on a base of 889 leisure visitors a year there is nothing to attribute against. Anyone quoting a return on generative-search spend for this territory is quoting an assumption.

On-island adoption of AI tools has not been measured. Neither has the volume of prompts about the territory, nor citation share for any Saint Helena source, nor how often AI summaries appear for queries about the place. Six categories, and the honest label for all six is unavailable.

Nor can the counterfactual be established. If arrivals rise, the causes plausibly include flight schedule changes, currency movement, a documentary, cruise itineraries, whale shark season and the marketing work, in some proportion nobody can decompose. The National Audit Office reviewed the £285 million airport investment in February 2025 and concluded the expected benefits had not yet been achieved. Even very large interventions here resist clean attribution.

What is left is still worth having. Whether the answers about the territory are accurate. Whether the errors are fewer this quarter than last. Whether the pages a machine would reach say what the government says. Whether enquiries, counted honestly, are more numerous over twelve months than over the twelve before. That is a real report. It is shorter and duller than the one most agencies produce, and it has the advantage of being true. The way this feeds into search work is set out on our SEO service page for Saint Helena, and the audit mechanics on the companion GEO page.

---

Frequently Asked Questions

Why should percentages be avoided when reporting on a market this small?

Because on a base of tens, a percentage measures noise. Eight sessions rising to twelve reports as a fifty percent increase, and the most likely explanation is one household using three devices. Twelve falling to six reports as a fifty percent collapse, and the most likely explanation is that the same household finished booking. Absolute counts over rolling twelve-month windows remove that distortion, and they match the way the St Helena Statistics Office reports arrivals.

Which visitor figure should a Saint Helena marketing report use?

For destination marketing, 889, the non-St Helenian leisure visitor count for the twelve months to October 2024. For accommodation demand, 1,983 all leisure arrivals is defensible because returning Saint Helenians also book rooms. Total arrivals of 5,298 belongs to logistics planning, not marketing. Whichever is chosen, name it at the top of the report and do not change it mid-engagement.

Can search volume for Saint Helena be measured at all?

Not by any mainstream tool. Ahrefs, Semrush and Google Keyword Planner apply minimum-volume thresholds and round small numbers toward zero, and a territory of fewer than four thousand residents never clears them. Google Trends is unusable for a different reason: it reports the share of queries rather than the count, and traffic attributed to this territory is heavily affected by VPN exit-node routing, which has put the territory at the top of worldwide interest tables for cryptocurrency.

Is a screenshot of an AI answer proof of visibility?

No. A January 2026 study by SparkToro with Gumshoe.ai, covering 2,961 runs of twelve prompts by around 600 volunteers, found the odds of the same list appearing twice were under one in a hundred, and the same list in the same order closer to one in a thousand. A screenshot is one draw from a wide distribution. The defensible alternative is a fixed prompt panel run repeatedly under declared conditions, reported as accurate runs out of total runs.

Why do published internet penetration figures for Saint Helena vary so much?

Because they measure different things at different times. DataReportal editions show 2,913 users at 55.2 percent in early 2024, 3,300 at 63.3 percent in January 2025, and 1,950 at 37.6 percent in October 2025, covering the combined territory rather than the island alone. The 2021 census, measuring households on the island, found 66 percent with internet access at home. Households and individuals are different units, island and combined territory are different geographies, and methodology changed between editions.

What can honestly be promised in an engagement like this?

Accuracy work and count-based reporting. Specifically: a defined prompt panel with logged outputs, a factual error register with a documented movement, corrections requested and secured on third-party pages, contradictions resolved within the owned corpus, and enquiry counts on a rolling twelve-month basis. What cannot be promised is a percentage lift, a share of voice figure, a citation guarantee, or an attributed return on spend, because none of those can be evidenced at this volume.

How often should reporting happen?

Less often than instinct suggests. Monthly reporting on a base of tens generates movement that has to be explained and did not mean anything, which erodes trust in both directions. Quarterly reporting with rolling twelve-month totals, plus an annotation log recording flight schedule changes, press coverage and seasonal events, gives a clearer picture and matches the planning cycle of a government-funded client.

When is the honest recommendation to stop?

When the error register is empty and staying empty, when answers about the territory are consistently accurate and cite official sources, and when the correction backlog on third-party pages has been cleared. At that point the problem being solved has been solved, and continuing generates activity rather than value. An engagement in a market this size should have a stated stopping condition from the outset.

---