A wine research firm called Wine Opinions ran a survey in January this year, and one number in it should have started a much bigger argument than it did. A quarter of American wine drinkers say they have used an AI tool to pick a wine, and then actually bought the wine it suggested.
Split that by age and it gets sharper. Among 21 to 39 year olds, 52% use AI for wine recommendations. Among 40 to 59 year olds it is 21%. Among the over-60s, 7%.
So the generation the entire wine trade says it cannot reach has found a way to get wine advice. It just is not asking anyone in the trade.
The reason is the part worth reading twice
Wine-Searcher went and asked working sommeliers about this in September, and one of them gave an answer so honest it is almost uncomfortable.
Judit Ayago’s explanation is that young guests turn to AI because they have inhibitions about wine caused by the formality that surrounds it, and a fear of making a mistake.
Read that again. They are not choosing AI because it is smarter. They are choosing it because it will not make them feel stupid.
Nobody has ever been sneered at by a chatbot. Nobody has ever pronounced something wrong in front of a language model and watched its eyebrows move. You can ask an AI what the difference is between Chablis and chardonnay, or whether a £9 bottle is any good, or what on earth “mineral” means, and it will simply answer. Ask the same three questions across a shop counter and you are taking a social risk, and quite a lot of people have learned the hard way that the risk is real.
We have been saying since 2018 that snobbery is not a harmless quirk of wine culture, that it is a commercial problem, and that it costs the industry drinkers. Here is that argument arriving as a market statistic, reported by the trade press, and sourced to a sommelier.
A number to handle carefully
The same reporting carried a French figure: 58% of 18 to 25 year olds in France have used an AI app to choose wine, and 41% of 26 to 35 year olds. It traces back to Le Figaro via Wine-Searcher.
We went looking for who commissioned it, how many people were surveyed and how they were asked, and we could not find any of it. Neither could the German trade title that summarised the same story. That does not make it wrong, but a percentage with no published method behind it is not evidence, it is a vibe with a decimal point.
The Wine Opinions numbers are the ones to lean on, because at least we know who ran the survey and when. A separate Dynata poll of a thousand American adults, from May last year, found 31% had already used AI to choose alcohol, which sits in roughly the same territory. The trend is real. The precise French figure, we would not quote without a caveat.
The fair objection
Not every sommelier in that piece was relaxed about it. Matthias Alberghi’s worry is that leaning on AI stops people thinking for themselves.
He has a point, and it is worth taking seriously rather than waving away. A machine that hands you an answer is not the same as learning why the answer is the answer. Wine is one of the few consumer goods where the pleasure genuinely deepens with understanding, and outsourcing every decision to a phone means you never build the map in your own head.
But we would put it differently. Somebody who asks an AI what to drink with roast chicken has already done the hard part, which is caring enough to ask. That person is not lost to wine. That person is at the very start of wine, and they went to the source that felt safe. The problem is not that they asked a machine. The problem is that the machine felt safer than us.
Now the part that should worry every estate
Here is what almost nobody in the trade has worked out yet.
If a meaningful share of drinkers are asking a model what to buy, then models have become a prescription channel. Not a marketing channel, a prescription channel, in exactly the way a sommelier’s recommendation or a merchant’s shelf-talker has always been. Something tells the drinker what to buy, and the drinker buys it.
For twenty years the wine trade has fought over the channels it could see. Critics, competitions, listings, shelf position, the wine list. A new one has opened, it is already carrying real purchases, and the overwhelming majority of estates have done precisely nothing about it, largely because nobody has told them it exists.
So what actually decides whether a model names your wine?
The research is younger and messier than the consultants selling “AI visibility” packages will admit, but two studies are worth knowing.
The first, from June this year, tested how established brands compete against newcomers inside GPT, Claude and Gemini. When two products had identical specifications, the famous brand was recommended 100% of the time. That is the bad news, and it is the news the consultants lead with.
The good news is in the next line. That dominance collapsed as soon as the challenger had any real advantage at all. Give the unknown product even a marginal quality edge and it took between 64% and 80% of the recommendations. The incumbent advantage is real, and it is thin. It holds only when there is genuinely nothing to choose between you.
The second study is a piece of work on LLMs as recommender systems by Lichtenberg, Buchholz and Schwöbel, and its finding runs against the panic. Compared with the traditional recommendation engines that have shaped what we all buy for the last fifteen years, the language-model recommender showed less popularity bias, not more. It was more willing to surface the obscure thing.
Both studies come with an honest health warning: one is about skincare, the other about films. Neither is about wine. We are reasoning by analogy here and we would rather say so.
What an estate should actually do about it
None of this requires an agency or a subscription. It requires writing down, in public, in plain words, what your wine is.
Describe the wine the way a drinker would. A model matches the language in a question to the language on a page. Someone types “light red for a hot evening, about fifteen euros”. If your site says “élevage of 12 months in 30% new oak, terroir of Kimmeridgian marl”, you have not answered that question in any language the machine can connect to it. Say it is light, say it is good cold, say it costs fifteen euros. Then say the Kimmeridgian part, for the people who want it.
Put the facts in text, not in a PDF and not in a picture. An enormous number of estates keep their grape percentages, their alcohol, their food matches and their current vintage inside a downloadable tech sheet or, worse, baked into a JPEG of a label. That information is functionally invisible. If a human cannot select it with a cursor, treat it as though it does not exist.
Exist somewhere other than your own website. These systems synthesise across sources. One page saying you are excellent is worth far less than a merchant listing, a regional body page, a press mention and a reference entry all independently describing the same estate in the same terms. Consistency across sources is doing more work here than eloquence on your own.
Spell your name the same way everywhere. Accents included. An estate that appears four different ways across the web is four weak entities rather than one strong one.
Then test it. This is free and takes ten minutes. Open two or three of the models, ask the questions a real customer would ask about your style, your region and your price bracket, and see whether you are in the answer. Most producers have never done this once. Whatever comes back is the most direct feedback on your public presence you will get this year.
Two warnings, because we are not going to pretend
The first is ethical, and the skincare study surfaced it plainly. The researchers found that invented authority language, clinical-sounding claims with nothing behind them, was worth roughly the same as a real quality improvement in the eyes of the model. That is a wide-open door, and wine already has more than enough people willing to walk through doors like that. Writing clearly about a wine you actually made is optimisation. Inventing distinctions to game a model is the same lie as a fake medal on a label, and it will age just as badly.
The second is strategic. The same study found that once everybody optimises, the individual gain collapses almost to nothing, while anyone who has not optimised gets recommended zero times. That is a prisoner’s dilemma, and it means this is not an opportunity so much as a floor. Doing it will not make you famous. Not doing it will make you absent.
The thing underneath all of it
Strip out the technology and the finding is old.
People are frightened of getting wine wrong. They have been told for decades that there is a right answer and that not knowing it says something about them. Given a source that would not judge them, half the under-40s went to it immediately and did not look back.
Being findable by a language model is worth doing, and any estate can do it in an afternoon. But it is the small version of the lesson. The big version is that the winning move, for a machine and for a human being standing in front of you, turns out to be exactly the same one: describe your wine in words an ordinary person can use, and do not make anyone feel small for asking.
