14,256 recorded answers · 10 to 25 August 2026
7 questions went to the category and 2 to each house by name. Every answer was archived on the day it was recorded, with its engine, its market and its date, and read against all 46 houses.
Every recorded answer receives one of five verdicts: 1 recommended, 2 named without endorsement, 3 named and discounted, 4 substituted, 5 absent.
The verdicts were named on 20 August 2026, before any figure was computed. No verdict is assigned by hand.
46 houses. Every eligible Swiss house exhibiting at Geneva Watch Days 2026. Houses under majority group control and non-Swiss houses are excluded. None was invited and none was set aside by judgement.
Comité Colbert and Bain & Company, June 2026. 35 houses and their clients.
The people who made their money with AI are the people who will ask an AI what to buy.
Reuters and BCG, 2025. Fédération de l'industrie horlogère suisse, 2025.
Similarweb, May 2026, via Momentic. Google's AI Overview is absent from this ranking because it is not an assistant a buyer opens. It is the answer above the results.
The link is no longer the channel. The citation is.
Chartbeat via Axios, March 2026. SEER Interactive, April 2026.
One house in nine is quoted at less than half its own price. Five are quoted above it.
Measured. Not published until two people have read it. The extraction rests on 588 answers, 4.4% of the set, and it mistakes a servicing cost for a price. Two readers are working through a sample of 200 separately. This chart publishes with their level of agreement printed beside it, or it does not publish.
By house the figure runs from
The individual sites, named, with the number of times each was read and the answers they fed, are in the source file.
The press falls 10.9 points. The marketplaces rise 11.3. They change places.
Share of everything cited, in answers that recommend against answers that discount. 7,804 sites in all. The largest single site is 3% of what the engines read, and past the first 30 there is nothing to buy a place on.
Two kinds of watch writing, and the machines treat them as if only one existed.
A house can be covered by every trade title in Switzerland and remain invisible to the machine.
Asked with no budget at all, Tissot comes first in 65.3% of answers. No house of the 46 is ever named first.
One independent house holds a first place at the highest level. It is not in the study and it does not exhibit at the fair.
Rolex is in almost every one of those answers. It is simply not the name the engine puts first.
Named first and named at all are two different things. When the buyer has not said what he can spend, the engine opens with the affordable name.
A house exists inside its own question and nowhere else.
1 of the 4 largest names appears in 62.1% of answers about an independent house.
For a house that has always said who makes its movements, the engine repeats what the house already publishes. For a house that has let the market believe it is vertically integrated, the engine now answers that question for the buyer, in front of him, without being asked.
Gemini and Google AI Overview carry no advertising anywhere.
One house of the 46 shares its name with a small car sold in Asia. The engine asks the buyer which one he means, then describes the car. Another shares its name with a sewing machine, another with an arms manufacturer.
16 of the 46 houses have their own name written incorrectly by an engine. Nobody told them.
Share of a house's own 288 answers. The houses are not identified here. Each is told its own figure.
One house loses the initial in front of its name in 248 of its 288 answers, so the machine calls it by half its own name. Another is written four different ways, two of which belong to a bakery group and a rum. A third is spelled as a company that makes non-lethal weapons, in 46 answers. A house that cannot be spelled cannot be found.
No causality. The record shows a movement. Attributing that movement to any action needs a controlled comparison, which does not exist before edition two.
No result from a single observation. Nothing seen once is presented as an engine's settled position.
No ranking of the 46 in any public document, under any name.
No engine reports which version of itself answered, and the model behind an answer can be changed by its publisher without notice.
One window, 3 markets, the English language.
The study does not measure what the buyer did next, and nothing in it says a watch was sold.
A figure that could not be computed honestly on the closed window is absent, not estimated.
AI engines are now where a wealthy buyer begins. That much is measured elsewhere and is not in dispute.
What this study measures is what happens next: the way an engine assembles an answer does not favour the house. It reads the market before it reads the maison, it reaches for the largest names when asked about the smallest, and it offers a discount to a buyer who did not ask for one.
For 200 years a house decided how it was described. It chose the words, the images, the shop, the salesperson and the moment. That is over. The description now happens somewhere else, before the house is involved, and for the first time the house is not in the room when it is written.
With the loss of control of the narrative of the brand, AI visibility becomes of strategic importance and will influence the goodwill of the house.
If a figure here concerns your house, or if you want to discuss what to do about it, write to Alexandre Olive.
The record cannot be recreated. The 16 days of August 2026 will not be run again.
No house of the 46 is named in this page.
The full method, the prompt set, the definition cards and the intervals are published at The Method.
The Luxury AI Report, Edition One: Swiss Watchmaking. USKALE SA, Genève, 31 August 2026.
Published figures may be quoted with attribution. The underlying record, the answers and the dataset are the property of USKALE SA and are not licensed for redistribution.