How the measurement is made, and what it cannot tell you.
Everything below is checkable.
Where we do not know something, this page says so.
What this page answers
What is measured
- What exactly is one reading?
- Can I have one number for “AI”?
- What question do you put to the assistants?
- What counts as a mention?
- Is a destination counted the way a business is?
- What do you show me about the other businesses in my niche?
- In which language is my market asked about?
- What exactly is behind a figure I see on a page?
How it is counted
What one reading is
- Questions
Fixed per niche and per market, in the language that market is asked in.
A question is never edited: a new wording starts a new series, and the two are never drawn as one line.
- Assistants
Perplexity, ChatGPT, Claude — each asked separately, at least once a week.
Never blended into one score, and the block below is why.
- Runs per reading
One run per assistant per question per morning.
- When we ask
The assistants are asked around 05:00 UTC; your site is read around 08:00 UTC.
A row carries a comparison from the fourth day.
4 days is the shortest span in which a move can be told from the ordinary movement of the row itself.
- A mention counts
only when the answer cites a source.
A name that appears without a citation is recorded and kept out of the figure.
- A day counts
only when every retained cell carried the same set of assistants.
Days that did not are left out of every comparison — 14 of the 80 judged so far were.
Why we never give you one number for “AI”
9.4%
Perplexity named the business
1,134 of 12,022 cells where all three answered
8.0%
ChatGPT named the business
964 of 12,022 cells where all three answered
7.0%
Claude named the business
837 of 12,022 cells where all three answered
1.3%
all three, same question, same morning
157 of 12,022 cells
When a business is named at all, 92.4% of the time it is named by some assistants and not the others.
Being visible is a property of a business and one assistant, not of “AI”.
A single blended score hides exactly the thing that is happening.
12,022 cells where all three answered the same question onthe same day, over 40 uniform mornings · 21 August 2026 to 18 October 2026 · read 20 October 2026
How much the answer moves on its own
Measured on a cohort that changed nothing about itself. This is the number you need before you believe any change is real.
| Reading | Moves by itself | Sample |
|---|---|---|
| One morning against the nextthe calendar’s next day, and the same number of questions on both | 11.3% flip the verdict | 2,071 pairs |
| A point reading one week against the lastone morning a week, which is what a weekly instrument takes | 26.3% changed at all | 506 pairs |
| The smoothed weekly leveltwo adjacent runs of seven comparable mornings, in points of its own share | 2.4 pts on average | 287 window pairs |
| A single morning against the whole periodits own reading · 21 August 2026 to 18 October 2026 · read 20 October 2026 | 12.9 pts average, 57.1 worst | 17 companies |
the research cohort, on the uniform days · 21 August 2026 to 18 October 2026 · read 20 October 2026
npx tsx scripts/public-figures.ts — howMuchItMovesByItself, over uniform days from day_instrument
This is why a figure here comes with a window and a band, and why we do not report one morning against the morning before it.
1. A question about the category, not about you
The first version asked the platform best travel companies similar to <name>.
The platform would go looking for that name, find the company’s own site and cite it.
That handed a mention to almost any business with a working website.
That measures whether a company has a website, not whether it gets found when someone asks about the category.
Those results are not used.
The question is now phrased as a search key — “best dive centres in Seychelles” — with no name inside it.
The wordings come from what people type into Google, read from Google Ads on 5 October 2026.
That is search, and we have measured that a key and a spoken question do not return the same businesses.
2. A mention only counts with a citation
A name matching somewhere in the text does not count.
Models routinely repeat the company name from the question in their own preamble with no source behind it.
Counting that means measuring the echo of our own query.
A mention requires a real citation: a source marker attached to the phrase where the company is named.
This bug reached working code three times — in the Claude, Perplexity and ChatGPT providers.
All three times it was caught only after a live answer was read by eye.
2a. Destinations are counted by a different rule
A business counts as mentioned only when the assistant names it and backs that sentence with a link to a source.
A name in prose is not enough.
Business names occur in ordinary sentences by accident.
We have twice found that counting them produced mentions that were not there.
That rule cannot be carried over to a destination.
When an assistant writes that the Seychelles are better for diving in October, it attaches no source to the word “Seychelles”.
A country is not a citable claim the way a business is.
Requiring a citation would measure nothing anywhere, in every language, for every destination equally.
So a destination is counted when the assistant names it in the text of the answer.
Where the answer also cites that destination’s own tourism board, we count that separately and report it separately.
It is a stricter reading of the same answer, never folded into the first number.
It is measured only on questions that name no destination at all: “best island in the Indian Ocean for diving”.
Not “Seychelles or Mauritius for a honeymoon”.
A question that names a destination guarantees that destination in the answer.
A share built on that would measure our own wording.
It is a share of answers, not a share of travellers. Nothing here says where anyone went.
A single run is not a measurement.
Asked five times in one day, the same question named one destination in three answers of three.
Another it named in between one and three.
Differences between single days are not read as change; the series is.
3. Two variances, not one ±
Within-day spread (platform non-determinism) and between-day spread (drift in the results themselves) are counted separately.
They are never collapsed into one number.
On the first day the spread reached ±37.7 and was called “model noise.”
Across three days, on one company, between-day spread stood at 42.3 against a within-day 15.1.
Whether that holds across the row is being measured.
Until it is, we do not present a step between two mornings as a change — only as a reading beside the spread it has to beat.
The practical consequence: a single-day reading cannot be presented as a measurement.
The same question, asked again on the same morning
On 27 September 2026 we put “where to learn to dive in Mauritius” to the three assistants.
We read the answer, and read it again the same morning.
One reading scored 0.
The other scored 100.
Of 404 cases where we read the same question twice on the same day, 44 came back different — 11%.
And 18 of those differed by fifty points or more.
Of the 65 businesses we measure in these markets, 45 are never named at all.
Their rows are flat because no answer has them in it, which is absence rather than stability.
Across the 726 day-to-day steps of the rest, the median change from one morning to the next is 8.3 points.
Those are two different measurements and both are ours.
0 and 100 is the spread WITHIN one morning, between two readings of the same question.
8.3 points is the move BETWEEN mornings, once a day is averaged.
A single reading is noise and a row of days is not.
So a day is not a measurement here.
We draw the spread beside every figure, and we do not call a one-day move a change.
Not in the dashboard, not in a letter, not when the move is upwards.
A fixed reading over the markets we may publish — not a live count.
The repeated readings were taken 7 October 2026, the day-to-day movement 24 October 2026.
What a move has to beat before we call it a change
12.1 percentage points: a company's own share minus its niche's over the same days, between two adjacent windows of 7 days.
It is the ninetieth percentile of the move companies made while nobody did anything.
Read over 28 window pairs and 2186 observations, 2026-11-08, by scripts/check-nine-threshold.ts.
11.7 percentage points: a company's own share minus its niche's over the same days, between two adjacent windows of 10 days.
It is the ninetieth percentile of the move companies made while nobody did anything.
Read over 22 window pairs and 1610 observations, 2026-11-08, by scripts/check-nine-threshold.ts.
Neither figure is a threshold for an individual business, and neither is applied to one.
They are readings over the whole cohort: the move companies made against their sector while nobody did anything.
What a business’s own page compares against is the day-to-day spread of its own row, printed beside the figure it belongs to.
What we show you about the others in your niche
Our measurement of the other businesses in your niche.
How often each is named, how that moves, and your own figure beside each of them.
Their figures, not their names — the names are in the answer itself, which every level sees.
A rank is the one thing this is not.
We print each business’s own figure beside yours, with the window each was taken over, and never a position in a list.
4. Comparable by engine count, not by date
The score is divided by the number of platforms that answered.
When one API key stopped working on 24 August, every niche ran on two engines.
That day’s average score came out higher than the day before, with no improvement in visibility whatsoever.
A mention by one platform out of two scores 50 instead of 33.
So days are selected by the number of engines that answered, recorded on every row, rather than by the calendar.
A day with an incomplete set is excluded entirely.
The query language decides whose market you can see
The language is chosen by measurement, not by a country’s official language.
Tested on four countries both ways: in Thailand the English query returned twice as many real operators as the Thai one.
In Brazil English returned zero and Portuguese returned nine.
What predicts it is not the country’s language but whether the industry faces outward.
Koh Tao and Phuket sell to foreigners and publish in English; Patagonia and the Brazilian coast sell to locals.
Had the rule been “translate everything non-English,” Thailand would have been measured noticeably worse.
What these measurements do not show
This section is mandatory and is never shortened.
The score is not a share of travellers.
It is the share of platforms with a confirmed mention, not the share of people who will see you.
Which page a traveller came through is not measured. The source map says where an assistant took a name from.
It does not say how anybody arrived, or whether a booking happened.
Changing a site has not been shown to change the naming.
Six checks across different data found no link between the state of a website and whether assistants name the business.
Each of the six is listed, with what it found, on the research page.
One phrasing is not a niche.
The same company scored 38, 100 and 67 on one question across three different days, with nothing changed on its side.
- The data shows absence, not its cause. Why a company is missing from an answer does not follow from these measurements.
- Position is not weighted. A mention on the first line and on the seventeenth line of an answer count the same.
Who else was named is read from the text, not from the score.
A second pass lists the businesses each answer names, in order.
A business’s place in that list is shown beside the score as an observation.
It is not weighted and not part of the score.
A name the model returns that does not occur in the answer is discarded.
On 606 stored answers, six were.
A count is a count of what the model said.
Our call to Claude caps its answer at 1,024 tokens, so its count covers what it said within that limit.
That is the same text a user would have seen, not the model’s whole answer.
Perplexity’s list of sources is not stored, so its count rests on the prose alone.
Gemini is excluded.
The Gemini API terms for Grounding with Google Search prohibit analysing results and using links to build an index.
That is exactly what this measurement is.
Perplexity, Claude and ChatGPT are measured.
Paying us does not change what we measure
What we measure does not depend on who pays us.
A business that becomes a customer is measured exactly as it was before.
No one can pay for a better result or a place in an answer.
Any business can ask to be removed from the measurement; that is free and is not tied to any purchase.
The figures on these pages, and what each is counted over
Each figure on the public pages carries, beside it, what it counts and over what window.
Whose rows it is narrowed to, and how to produce it again, are here — one entry per figure, from the same record the page prints.
- Answers that named at least one business
- the thirty days to 14 October 2026 · the markets we may publish, on the businesses in our research cohort, asked in the search-key form the series uses · read 14 October 2026
- npx tsx scripts/public-figures.ts — scan_results joined to companies with source = manual_research and a publishable country, counted once per (morning, assistant, question), distinct business names per answer
- Mentions — a business named in an answer, counted once for each answer that names it
- the thirty days to 5 November 2026 · the markets we may publish, on the businesses in our research cohort · read 5 November 2026
- npx tsx scripts/public-figures.ts — the same answers as NAMES_PER_ANSWER, counted by distinct business rather than by answer: every name in every answer is one mention, and the shares are of that total
- The share of answers that named somebody from an industry
- the twenty usable days to 31 October 2026 · Seychelles, the Russian series of the same questions. Shares are over the businesses we measure in each industry, never over one company; an industry with fewer than 5 businesses measured is listed without a share · read 7 November 2026
- npx tsx scripts/sector-table.ts --language=ru --period=longest chooses the window and gives the named shares and the company counts; npx tsx scripts/industry-table.ts --market=Seychelles --from=<start> --to=<end> gives the citation shares OVER THAT SAME WINDOW — the dates are copied from the first, never defaulted, or the record holds two windows
What has changed in the method itself
The counting rule changed on 2 September 2026.
Previously, in an enumeration naming several companies under a single citation, only the last one named was credited.
The error was one-directional — it understated visibility — and it lived in two engines out of three.
Measurements before and after that date are not comparable with each other.
We write this here because any chart crossing 2 September will show a step, and it is easy to mistake for a rise in visibility.
A second limit has the same shape and a different date.
The full text of each answer has been stored only since early September 2026 — the first stored answers are dated 4 September.
Before that we kept whether a business was mentioned, and nothing else.
Three dates appear across this site, and they are not versions of one date:
- 21 August 2026 — the first market’s daily series begins. This is the oldest day any row of days can reach back to.
1 September 2026 — the window the published niche comparison is drawn from.
It starts later than the series on purpose.
It is a period chosen to be the same for every market in it, and not every market was being measured in August.
4 September 2026 — the first answer whose full text was kept.
Anything built on reading an answer rather than scoring it — who else was named, which sources were cited — can only begin here.
Never earlier.
Nothing reconciles them, because there is nothing to reconcile.
A measurement can be older than the text it was read from, and a published comparison can be narrower than both.
Which other businesses were named, and which sources were cited, is read back out of the answers afterwards.
So that too begins in September.
It cannot be recovered for the days before it.
The daily series itself runs from 21 August, so its first two weeks carry visibility and no names.
Any list of who else was named will start in September however late it is produced.
That is a gap in what was kept, not a change in what the assistants said.
What every reading is made of
What we ask
Real traveller phrasings, from a table built and tested for each market — never a question with your name in it.
Who we ask
The same question, word for word, to Perplexity, Claude and ChatGPT.
What counts as a mention
Only a name the answer’s own source backs. A name with nothing behind it is not counted.