The method
A complete population of Swiss watch houses, measured in full.
The population, the questions, the engines, the markets, the sampling and the rule that assigns the verdicts.
The Luxury AI Report is published by ELIA Atelier and measures how AI answer engines describe the luxury industry. Edition one, published 27 August 2026, records 17,820 answers across 46 Swiss watch houses, nine buyer questions, four engines and three markets.
How this differs from the other readings
| The Luxury AI Report | Readings of this kind, in general | |
|---|---|---|
| Where the figures come from | Recorded answers. The questions were put to the engines and the replies written down. | Commonly modelled estimates derived from public signals, domain authority and retrieval patterns. |
| Population | A complete population. 46 of 46. | Commonly the largest names, chosen by size. |
| Engines | Four, named, with model versions recorded. | Varies, and model versions are rarely stated. |
| Markets | Three. United States, United Kingdom, Singapore. | Usually the United States alone, in English. |
| The rule | Fixed, dated and published before sampling. | Often undisclosed, or applied after the fact. |
| What is measured | Five qualitative verdicts, including whether the engine steers the buyer to a lower price. | Usually a citation share. |
| The measured house receives | Its private letter, without charge, whether or not it ever replies. | Often nothing, or a paid audit on request. |
| Can it be checked | The questions, the rule, the window and the arithmetic are published, so the measurement can be rerun. | Rarely a published protocol. |
A citation share can be estimated from how often a name appears in the sources an engine is likely to read. What an engine tells a buyer, in a given market, on a given day, can only be found by asking it.
Nobody had asked.
The population
A house too small for any top twenty-five appears here, so the whole distribution is visible rather than its summit.
The edition also reports every house the engines named, whether or not it belongs to the forty-six. No comparison set was chosen in advance. The field of names is whatever the engines produced, counted from the recorded answers.
The eligibility rule
Swiss watch houses on the official Geneva Watch Days 2026 exhibitor list. No other exclusion.
Houses majority-owned by LVMH, Richemont, Swatch Group, Kering, Chanel or Rolex.
Houses that are not Swiss, whether or not their watches are Swiss made.
Minority stakes held by any group, including Chanel, kept as an attribute and never used to exclude.
The rule yields forty-six houses for edition one, and it can be rebuilt by anyone working from the public exhibitor list.
The question is which houses never appear in the answer at all.
Markets, window, volume
Hong Kong was the intended third market and was withdrawn before sampling began: ChatGPT is not available there, so a house's answers could not be compared across all engines in that market. Singapore, the second largest watch market in Asia, replaced it. Any change is announced before sampling begins.
Nine buyer questions, each put eleven times, make ninety-nine prompts.
- Prompts per market, per engine, per day
- 99
- Engines
- 4
- Markets
- 3
- Consecutive days
- 15
The window ran from [START DATE] to [END DATE] 2026. Model versions and access dates are recorded for every run.
Nine buyer questions, each put eleven times, give ninety-nine prompts in every market, on every engine, every day. Four engines, three markets, fifteen consecutive days: 17,820 recorded answers. English throughout. Every answer is read against all forty-six houses.
Every answer is stored with its date, market, engine, model version and full text. Model versions and access dates are recorded for every run and stated with the figures, because a figure quoted without a model version cannot be interpreted.
Fifteen consecutive days is a chosen window: research published this year at the University of St. Gallen found that only around 35 per cent of the sources an engine cites persist from one day to the next, so a single screenshot of an AI answer is noise. The Report records across a fifteen-day window.
The nine questions
Nine questions, written the way a buyer puts them. The full wording of all nine publishes with the edition.
- Three concern the gift: the occasion, the recipient, the budget.
- Three concern the purchase for oneself: a stated taste, a complication, a case size.
- Three concern the house itself: what it is known for, how it compares with a named peer, and whether it holds its value.
Wording is held constant across markets and engines for the full window.
The five verdicts
Each answer receives one of five verdicts.
The five definitions are on the front page. The rule that assigns them is in section 07.
The rule, and what is published
The rule was written, dated and published before the first question was asked. No verdict is decided after the fact.
Every answer is read against that rule. A sample is then read again by hand, and the two readings are compared, so the coding can be checked rather than trusted.
The nine questions, the rule, the engines, the markets and the window. Anyone who doubts a figure can put the same questions to the same engines and see what comes back.
The answers themselves. Each measured house may obtain its own complete record at the published price. The complete record for a category stays with the practice.
The Report reports what the engines said during a fixed window. It does not predict what they will say next, it does not rank houses by quality, and it offers no investment guidance.
Every figure it publishes can be checked.
Independence
No house can pay for inclusion, for placement, or for removal.
The population and the rule were fixed and published before sampling began.
Every measured house receives its private letter without charge.
The Report is published by ELIA Atelier, a practice that advises luxury houses on this subject and is paid to do so. That is stated here so a reader can weigh it.
Any house that is both measured and a client of the practice is named in the edition, and no engagement of any kind changes how a house is measured: same roster, same days, same rule, same verdicts.
The Private Dossier is priced at CHF 2,000 for every measured house, published here, identical for the smallest atelier and the largest manufacture. The price is not negotiated, for the same reason it accepts no sponsorship: a measurement that can be bargained with can be bought.
Method enquiries, including requests to see the verdict rule or the recorded model versions, to alexandre@eliaatelier.ch.