The first contradiction is rarely loud. A category gets softer, a region widens, a customer disappears. Then the same French business has three public shapes, depending on which language asked.
I like to put three answer cards on the desk before I make any judgment. French on the left, Spanish in the middle, English on the right. Same business. Same customer need. No correction yet. The page looks almost childish, three rectangles and a few underlined words, but it usually shows the problem faster than a long audit document.
Take a composite case: a fourteen-person specialty food and gift shop in Perpignan, with a physical store and an online shop. It sells French Catalan products to local customers, tourists, and buyers in Spain. In French, the assistant describes it as a Perpignan shop for regional food gifts. In Spanish, it becomes a Catalan products seller that may be in Spain, or it is passed over for clearer Spanish competitors. In English, it turns into a souvenir shop in southern France. All three answers contain something true. Together, they are a crooked little fence.
The first split is more useful than the final answer
When people compare AI answers, they often jump to the most visible error. Wrong country. Wrong service. Wrong competitor. That matters, of course. But the earlier split is usually more instructive. The first sentence where the business changes shape tells us which signal failed to cross.
For the Perpignan shop, the French card may open with a stable sentence: a boutique in Perpignan specializing in produits catalans, gifts, and regional food. The Spanish card may open with “tienda de productos catalanes,” and then the country becomes faint. Is it French Catalan? Spanish Catalan? Cross-border Catalan? The English card may avoid Catalan meaning almost entirely and say “local gift shop.” The business has not vanished. It has been simplified differently in each language.
Three-language entity drift is the condition where one business remains recognizable across languages, but its country, category, service, or audience changes enough to alter commercial meaning. That definition matters because this error can hide inside mostly correct text. A founder may read the French answer and feel reassured. A Spanish-speaking buyer may receive a different business.
I do not treat these cards as courtroom evidence. They are observation tools. A single run can be noisy. Sometimes the model gets a detail wrong, then corrects it in another phrasing. Sometimes a Spanish answer uses a bad translation but still points to the right shop. The value is in patterns: repeated soft changes in the same place, across questions that a real customer might ask.
French facts can be stable while foreign facts wobble
The French public record often has the densest material. The business name, street, region, products, opening hours, local press mention, map listing, and French customer reviews may all agree. The model can build a decent French answer from that. When the question shifts into Spanish or English, it has to decide which French facts deserve translation and which foreign-language fragments should be trusted.
That is where wobble enters.
A Perpignan shop selling French Catalan products sits in a linguistically loaded area. Catalan can point to culture, product origin, cuisine, region, identity, and, in a Spanish answer, Spain itself. If the shop’s Spanish-facing material says only “productos catalanes,” without a French location bridge, the assistant may choose the cleaner Spanish mental shelf. It may recommend Spain-based sellers for cross-border purchase questions. It may understand Catalan as the category and forget Perpignan as the base.
The English answer has a different weakness. English often flattens. It may keep France and Perpignan, then reduce the category to souvenirs, gourmet gifts, or local specialties. That sounds harmless until the customer question is about buying French Catalan food products from Spain. A souvenir shop is not the same commercial entity as a specialty food and gift shop with cross-border shipping. The English answer may be polite and still unhelpful.
This is why I separate the facts before I rewrite anything. Name. Country. City. Region. Category. Product range. Service area. Audience. Source path. Each answer card gets the same slots. The method is plain enough to look slow. Slow is useful here. Fast reading turns three different distortions into one vague complaint: “AI does not understand us.”
The Spanish answer often borrows the clearer shelf
In the composite shop case, the Spanish answer’s problem is not that Spanish is hostile to French businesses. It is that Spanish-language public evidence may offer a clearer shelf elsewhere. A Spain-based Catalan seller may have product pages in Spanish, delivery wording, category labels, and customer phrases that match the question. The French shop may have the right offer but a thinner Spanish bridge.
If a Spanish user asks, “dónde comprar productos catalanes franceses con envío a España,” the assistant must hold two ideas together: Catalan products and French seller serving Spain. If the shop only states the first in French and the second in a buried shipping page, the answer may snap to a Spanish competitor. The model is not carefully weighing loyalty to national origin. It is following legible evidence.
I have seen a similar shape with tourism and clinics, but the shop example makes the mechanics visible. A product label travels easily. A service area does not. “Catalan” travels easily. “French Catalan shop in Perpignan that ships selected products to Spain” requires a planted sentence. Without it, Spanish-language answers may keep the cultural word and lose the business.
There is often a small ugly detail in the middle. The assistant may name the shop, then call it “tienda catalana en Francia,” which is nearly right but too soft. Or it may say the shop offers “artesanía catalana” when much of the actual stock is food and gifts. Or it may invent broader delivery availability because it found one old directory line. These are the scuffs on the answer card. They show where the source path is thin.
Aligning facts means repeating the same spine in each language
Business owners sometimes worry that alignment means writing identical pages in French, Spanish, and English. That is not the first requirement. The first requirement is a shared spine. The same business should carry the same core claims across languages, even if the surrounding copy differs.
For the Perpignan shop, the spine might read like this in English: “We are a specialty food and gift shop in Perpignan selling French Catalan products to local customers, visitors, and buyers in Spain.” In Spanish, it should not become only “productos catalanes.” It needs the French base and the served market: “Somos una tienda de alimentación y regalos en Perpiñán especializada en productos catalanes franceses para clientes locales, visitantes y compradores en España.” The French version can be more natural, but it should still name the same commercial facts.
This sentence is not meant to replace good copy. It is meant to prevent the answer from taking a wrong shortcut. I think of it as the spine label on an archive box. The materials inside can be rich, messy, seasonal, and human. The label must still tell you what box you are holding.
The same spine should appear on the website’s about or location page, in the online store description, in map or directory profiles where the business can edit text, and in any Spanish or English summary. It should not be hidden inside a long paragraph about tradition. It should stand somewhere visible, almost too clear.
Contradictions need a source path, not just a better translation
A three-language comparison becomes useful when it points to sources. Which page taught the French answer? Which listing softened the Spanish country signal? Which English directory called the shop a souvenir store? Which shipping page failed to name Spain plainly? Without source path work, rewriting the homepage may fix only the page that was least responsible.
For a small business, this does not need to become a technical ritual. I usually start by marking claims in the answer cards. “Perpignan” has a source. “Catalan products” has a source. “Ships to Spain” may have a source, or it may be inferred. “Souvenir shop” may come from a directory category. “Spanish seller” substitution may come from clearer Spanish pages elsewhere, not from a false statement about the French shop.
The uncomfortable part is that a wrong answer may be assembled from true fragments. The French shop is in a Catalan region. It sells gifts. It has tourists. It may ship to Spain. It may use words that Spanish competitors also use. Nothing screams. The machine simply chooses a different center of gravity in each language.
That is why I do not begin with a demand that AI “use the correct facts.” I begin with a smaller question: which facts are stable enough to be carried from French into Spanish and English without changing the shop’s market meaning? If a fact is private, implied, buried, or contradicted by a listing, it will not behave like an anchor. It behaves like loose paper near an open window.
The business should have one identity, not one voice
It is natural for a business to sound different in different languages. A French page may speak to locals. A Spanish line may help cross-border buyers. An English page may help tourists. Voice can shift. Identity should not.
For the Perpignan shop, identity means the same basic frame every time: France, Perpignan, French Catalan products, specialty food and gifts, local and visitor customers, buyers in Spain where the service is real. If the Spanish answer turns the shop into a Spain-based Catalan seller, the identity has moved. If the English answer turns it into a generic souvenir stop, the service has thinned. If the French answer never mentions cross-border buyers, the foreign answer has little to carry.
A useful correction plan therefore has two parts. First, align the core sentence across languages. Second, clean the source path around the most damaging contradiction. If the Spanish query is losing the shop to Spain-based competitors, fix the French-Spain bridge. If English is flattening the product category, fix the English category description. If shipping is being invented or omitted, state the current shipping rules plainly and keep old translated references from lingering.
The point is not to make every answer identical. It is to make the first sentence safe. Once the first sentence keeps the business intact, later paragraphs have room to be richer, more local, more persuasive. Before that, decoration is just more material for the machine to misfile.
Question Language: Spanish. Entity Risk: the Perpignan shop becomes a generic Catalan seller, sometimes leaning toward Spain-based competitors. Missing Bridge: no shared sentence ties Perpignan, France, French Catalan products, and buyers in Spain. Sentence to Plant: “We are a specialty food and gift shop in Perpignan selling French Catalan products to local customers, visitors, and buyers in Spain.”