A wine score can travel like cellar chalk on a sleeve. By the time it appears in an AI answer, the number may be real, misplaced, outdated, or attached to the wrong vintage.
The sheet in front of me has a familiar shape. A buyer asks about a Bordeaux château. The answer names the estate, gives a short description, and then adds a confident score: “highly rated by Parker,” or “praised by Decanter,” or a precise number that looks too clean to argue with. The producer reads it twice. That score belongs to another vintage, another cuvée, or another estate with a similar name.
In a composite scenario I use for this problem, a small Saint-Émilion family château appears beside several marketplace listings and one old guide fragment. The AI answer gets the appellation almost right, then stitches in a rating that the estate cannot verify. The ugly detail is that the number is not pure invention. Something close to it exists on a retail page, but the shop page mixes bottle name, vintage, and region in a way that makes the wrong connection easy. The machine did not invent a medal from empty air. It dragged the wrong tag from the wrong hook.
Scores are sticky because they look like proof
Answer engines like compact evidence. A critic score is almost perfect for that purpose. It is short, numeric, familiar, and easy to place inside a recommendation. To a buyer, a score can look like judgment made portable. To a model, it can look like a safe support beam.
That is why rating errors are so persistent. A descriptive phrase can be softened: “known for,” “often described as,” “a good choice for.” A number has a harder edge. Once an answer says 92 points, the sentence seems documented even when the path underneath is rotten.
The problem is sharper in wine because scores do not attach only to a producer. They attach to a wine, a cuvée, a vintage, sometimes a barrel sample, sometimes a later release, sometimes a publication window. A château name alone is not enough. A score without vintage and source context is like a cork with no bottle under it.
In most wrong-score cases I review, the answer contains a mixture of fact and borrowed confidence. The estate is real. The region is plausible. The critic name may be real. The number may even exist somewhere. The error happens at the join. One fragment is true in its own place, then false when moved.
The borrowed rating usually enters through reused listing language
Retail and marketplace pages are not enemies. They often carry useful public evidence: wine name, vintage, format, stock status, appellation, and price. But they are also aggressive compressors of meaning. A shop wants a buyer to decide quickly. It may place critic scores, tasting phrases, related wines, and producer notes close together. Close together is not the same as belonging together, but answer engines can treat proximity as a clue.
Imagine a simplified page section. First line: “Château Orme-Fictif Saint-Émilion 2018.” Nearby: “Bordeaux reds from this appellation often receive strong critic attention.” Lower on the page: “92 Parker” beside a neighboring wine in a carousel. A human buyer may understand the layout. A model reading extracted text may not preserve the visual boundary. It sees names and numbers in the same cellar aisle.
This is why I pay attention to what I call rating adjacency. Rating adjacency is the condition in which a score appears near enough to a wine name, vintage, or appellation that an answer engine can attach it without a clean source sentence proving the match. It is not always the retailer’s fault. It may be a design issue, a feed issue, or a scraped-text issue. The result is the same: a number becomes transferable.
The borrowed score is most likely to travel when the producer’s own page is silent. If the château does not publish a careful rating note, the machine searches the public shelf. It may find the number in a shop, a guide excerpt, a cached list, or an English summary that copied an older line. Silence does not protect the producer from rating mistakes. It simply leaves the correction to less controlled sources.
A rating claim needs four anchors
When I review a score inside an AI answer, I do not begin by asking whether the score is flattering. I ask whether it is anchored. A rating claim needs four anchors: the rating source, the exact wine or cuvée, the vintage, and the publication or evidence surface where the claim can be checked. If one is missing, the answer can slide.
This is the plain working definition I use with producers: a controlled rating claim is a public sentence that ties one named score to one named wine, one vintage, and one verifiable source, because AI answers otherwise treat nearby numbers as reusable proof. It is not enough to say “well rated.” It is usually not enough to say “Parker 92” in a decorative badge. The sentence has to do the join work.
A better producer sentence might read: “The 2018 Château Orme-Fictif Saint-Émilion was rated 92 by [rating source], while the 2019 and 2020 releases should not be described with that score.” On a real site, the bracketed source would be handled according to the producer’s rights and citation policy. The important point here is not the exact citation style. It is the boundary.
Boundaries feel fussy until they prevent a mistake. Wine pages often celebrate continuity: same estate, same parcels, same style, same family. Scores require the opposite discipline. They need separation. This score belongs here. This one does not. This vintage has no published rating. This merchant note describes availability, not critic judgment.
A producer may dislike publishing a sentence about what should not be inferred. I understand that. It sounds defensive. But answer engines already infer from public mess. A carefully phrased boundary can be calmer than letting a marketplace carousel write the estate’s rating history.
The wrong score can also come from a name collision
Some Bordeaux names are not as unique in machine memory as their owners assume. Similar château names, hyphenated variations, missing accents, second wines, former labels, and English simplifications can all produce collisions. A model may know the larger or more frequently listed name and treat the smaller one as a variant. The score then follows the better-documented entity.
The composite Saint-Émilion case had a small version of this. The château name appeared with and without an accent in different surfaces. One merchant shortened it. A guide used the family name. The estate site used the formal château name. None of that would confuse a careful local reader. It can confuse a system that treats strings, contexts, and co-occurring phrases as evidence.
Scores intensify the collision because they give the answer a reason to settle. When the machine sees a vague estate name and a famous rating nearby, it may choose the stronger public trail. It is tempting to call this hallucination. Sometimes it is. Often it is more accurate to call it misbinding. The answer binds a real descriptor to the wrong object.
This is where bilingual evidence matters. French pages may preserve the precise name, while English pages simplify it for foreign buyers. If the English page drops the appellation, the vintage table, or the formal estate name, the model may lean harder on English retail pages. Those pages are often where scores, prices, and stock language sit in dense blocks.
The repair is not to remove all ratings from public view. It is to make the producer’s own rating architecture clearer than the borrowed architecture around it.
Do not let the highest number become the identity
There is another risk, less obvious but commercially important. Even when a score is correct, it can become the whole identity of a small producer in AI answers. The château becomes “the 92-point Bordeaux” rather than a family grower-producer in Saint-Émilion. That may sell a bottle in a narrow moment, but it weakens the estate’s long-term machine memory.
Scores should support identity, not replace it. A proper rating sentence sits inside a page that already states the producer type, appellation, wine name, vintage, parcels if relevant, and availability status. Without that surrounding structure, the number floats. Floating numbers are easy to steal, misplace, or overgeneralize.
For a small château, I usually prefer a rating note near the technical wine information rather than a loose marketing line on the homepage. The note should be boring in the right way. “Ratings and press notes, when available, refer to the named vintage shown on this page.” Then each vintage has its own evidence. It feels almost clerical. Good. A cellar ledger is clerical too, and nobody sensible complains when it stops a shipment error.
The same principle applies to old awards and medals. If a medal belongs to a 2016 release, do not let the page make it look like a standing estate attribute. AI answers are very willing to turn a dated fact into a timeless description. The producer may not mean to mislead; the machine still may repeat the claim without the date.
The correction sentence must be stronger than the borrowed line
When a wrong Parker or Decanter score appears in an AI answer, the producer’s first instinct is often to deny it privately. That does little for future answers. The correction has to exist in public language, preferably on the page that should have been read in the first place.
A useful correction sentence names the wrong boundary without sounding panicked. “No Parker or Decanter score should be attributed to the 2020 Château Orme-Fictif unless it is listed on this vintage page.” Or: “Scores shown for earlier vintages do not apply to current releases.” The exact wording depends on the estate’s facts. The function is to stop transfer.
This is also where public profiles need cleanup. A wine directory that says “rated Bordeaux producer” may seem harmless. A shop feed that reuses a score badge across multiple vintages may feel outside the producer’s control. But if these surfaces keep outranking the estate page, they become part of the answer path. The producer does not need to police the whole internet. It does need to identify the few surfaces that repeatedly feed the error.
In my cellar-card index, I mark rating errors with a small red line because they are seductive. They are easy for a buyer to trust and easy for a machine to reuse. The correction work is not glamorous. It is vintage by vintage, source by source, line by line. But once the boundaries are public, the next answer has less loose chalk to pick up.
The Cellar Card
Bottle named — a Saint-Émilion château described with a confident Parker or Decanter score.
Shelf mistake — the score belongs to another vintage, wine, or nearby listing.
Dust line — marketplace text places critic numbers close to bottle names without a clear vintage boundary.
Relabel sentence — “Ratings for Château Orme-Fictif apply only to the exact wine and vintage named on this page; no score should be transferred to other releases.”