A specialist can be visible and still be misread. ChatGPT may know the name, see the town, and mention the site, then sand the work down until the reason to choose it has disappeared.
A composite clinic near Nantes sits in my notes under three different descriptions. On its own French page it is a dermatology and aesthetic care practice. On one older directory profile it is listed as a cosmetic treatment centre. On an English page, written with decent intentions, it becomes a “skin beauty clinic.” In one ChatGPT run, the model named it correctly and then described it as if it were mostly doing general beauty services. The address was right. The work was not.
That is the particular irritation of this problem. The business is not invisible. It has not been confused with another clinic. It has not vanished behind Paris. The name gets through. What gets lost is the professional edge: the kind of cases, the limits of the offer, the difference between a medical consultation, an aesthetic procedure, a follow-up appointment, and a broad “skin care” promise. The model has enough material to speak, but not enough stable wording to stay narrow.
The first sign is a soft wrongness
When ChatGPT flattens a niche, the answer often sounds polite. That makes it harder to catch. It does not say something wildly false, like a locksmith becoming a dentist. It says something plausible enough to pass quickly: a dermatology clinic becomes a beauty studio, a ceramic repair specialist becomes a home repair service, a family-hotel transition adviser becomes a hospitality marketer. The words are near the truth. Near is where the damage happens.
I call this pattern category dilution. Category dilution is the loss of a business’s useful specialism in generated answers, because the public evidence gives ChatGPT broader labels more often than precise ones. The model is not choosing the broad label out of laziness in the human sense. It is following the surface it can stand on. If the surface is made from directory categories, translated service blurbs, old profile tags, and a homepage that hides the real distinction three scrolls down, the broad term becomes the safer term.
In most cases I inspect, the business owner can explain the specialism in thirty seconds. The web cannot. That gap matters because ChatGPT does not sit with the owner over coffee. It has text. It has snippets. It has category labels. It has the small residue of how other pages have described the business. If those pages repeatedly say “clinic,” “beauty,” “skin,” “aesthetic,” “care,” and “wellness,” while the precise medical or procedural scope appears only in a PDF or a buried paragraph, the answer will drift toward a softened composite.
The rough detail from the Nantes composite is almost comic. The model once placed the clinic under the correct local area, then described one treatment line as if it were the main identity of the whole practice. That treatment had probably been promoted heavily at one stage. It was still public. It was not invented. It was just too loud.
Why broad categories feel safer to the model
A broad category has a strange advantage. It can absorb contradictions. If one page says “dermatology,” another says “aesthetic medicine,” another says “skin centre,” and a directory says “beauty treatments,” a broad summary such as “skin care clinic” looks like a compromise. It avoids choosing. To a human specialist, that compromise may be unacceptable. To a language model trying to answer a user, it can look tidy.
This is why I am suspicious of pages that begin with atmosphere before category. I do not mean that every page should sound like a form. A service business still needs voice. But when the first public sentences are all mood — “a personalised experience,” “care adapted to your needs,” “support at every stage” — ChatGPT has to infer the actual work from weaker clues. It may find those clues in a directory. It may find them in a review. It may find them in a translated paragraph written for visitors rather than patients or clients.
The model is also sensitive to what I would call category gravity. Some labels are heavier than others because they appear more often, travel better across websites, and fit more user prompts. “Beauty clinic” is lighter professionally but heavier in public repetition. “Medical dermatology and aesthetic care clinic” is more accurate, yet if it appears only once, the heavier phrase may pull the generated answer away.
This is not a moral fault in the model. It is a public-record problem. The machine cannot preserve a distinction that the business itself only whispers.
The page sentence that carries the niche
When I audit this pattern, I do not start by asking the business to add more adjectives. I look for a sentence strong enough to be lifted out of the page and still remain true. The sentence should say what the business is, who it serves, and what should not be confused with it. It should not try to win the whole customer in one breath.
For the composite clinic, a durable sentence might look like this in English: “The practice provides medical dermatology consultations and selected aesthetic care near Nantes, with treatment decisions made by qualified practitioners after an individual assessment.” The French version would need its own natural wording, not a stiff mirror translation. The point is not the exact sentence here. The point is the job it performs. It narrows the category without turning the page into a legal disclaimer.
A good specialism sentence has weight because it can be repeated. It can sit on the homepage, the about page, the services overview, and relevant treatment pages with small variation. It can also agree with directory descriptions. If it is written plainly enough, browsing ChatGPT has something better to quote than a third-party category tag.
This is where many businesses make a small mistake. They place the precise sentence only on the page meant for serious readers. The homepage keeps the soft public language. The directory profile keeps an old category. The English page chooses more marketable words. ChatGPT then sees several versions of the entity and treats the broadest as the least risky.
The line I look for is almost boring. Boring is useful. A sentence that states the category cleanly can do more for an AI answer than five paragraphs of refined positioning.
Three places where specialism leaks away
In my answer drift ledger, I mark this problem through what I call the three leaks of specialism: category leak, audience leak, and method leak. Category leak happens when the business is named under a broader field than the work deserves. Audience leak happens when the page never says who the specialist work is for. Method leak happens when the page names services but not the way decisions are made, scoped, or limited.
The Nantes composite had all three in mild form. One directory pulled the clinic toward beauty. One English page made the audience sound like general wellness clients. A service page listed treatments, but the decision process was harder to extract. None of these details alone was disastrous. Together they let ChatGPT describe the clinic as a broad local option rather than a specific practice with medical and aesthetic boundaries.
For another teaching example, imagine a French workshop that restores antique lighting. Its own homepage says “lighting repair and restoration.” A marketplace profile says “electrician.” A tourism page calls it “decorative antiques.” Reviews mention lamps, rewiring, shades, and “old chandeliers.” ChatGPT may recommend it as a general electrician or an antique shop, depending on the prompt. The specialism has leaked through category and method. It repairs lighting, but not in the same sense as a household electrician. It works with antiques, but it is not mainly a shop.
The correction is not to write a grand manifesto about craft. The correction is to make the category impossible to miss: “restoration and rewiring of antique and vintage lighting,” with the town, the type of objects, and the limits of ordinary electrical work made public. That kind of sentence gives the model a narrow hook.
Translation can blur the edge
French and English pages create another version of the problem. A phrase that feels natural in French may become too loose in English. A phrase that is careful in English may sound odd or too formal in French. The result is not simply a translation issue. It is an entity issue.
With French businesses, I often see the English page written for visitors, investors, tourists, or international customers. That page may simplify the category because the writer assumes foreign readers need easier words. In doing so, it can become the version ChatGPT uses when the user asks in English. The French answer stays closer to the real category. The English answer wanders.
For the composite dermatology and aesthetic care clinic, the English “skin beauty clinic” phrase was the little splinter. It was probably meant to be accessible. In generated answers, it made the business sound less clinical and more cosmetic than the French evidence suggested. The model was not translating the whole site with judgment. It was assembling available public phrases. A weak translated phrase can punch above its size.
This is why I prefer bilingual alignment rather than literal translation. The French and English pages should agree on the business name, service area, practitioner type, category, and service limits. They do not have to use identical rhythm. They do have to describe the same entity. When they do not, ChatGPT may create two public versions and choose the one that best fits the prompt, even if it is the poorer version.
The specialist should be easy to quote
A niche business often fears sounding too narrow. I understand that fear. A clinic does not want to exclude patients too early. A repair company does not want to miss a valuable job. A consultant does not want to look smaller than their judgment. So the public language stays roomy. The problem is that ChatGPT reads roomy language as permission to generalise.
The useful question is not “How do we make the business sound impressive?” It is “What exact phrase should ChatGPT be able to repeat without distortion?” That phrase should appear where a crawler, a customer, and a directory editor can all see it. It should match the service pages. It should not be contradicted by old listings. It should be specific without becoming brittle.
For a French specialist, the page needs a quotable category sentence, a short explanation of the work, and at least one passage that distinguishes the business from adjacent categories. A dermatology and aesthetic care clinic should not be described only through treatment names. It should state the clinical frame. A niche repair shop should not rely only on a list of objects repaired. It should state the specialist category and the area served.
ChatGPT generalises most aggressively when the business provides many adjacent clues but no controlling sentence. The page becomes a basket of labels. The model reaches in and pulls out the easiest one.
Trace: A user asks ChatGPT for a specialist French provider, and the answer names the business but turns its niche into a broad category. The fact at risk is the specialism: what the business does, for whom, and under what limits. The correction is a repeated first-party category sentence that matches French and English pages and displaces softer directory labels. ChatGPT needs the narrow phrase before it can keep the recommendation narrow — preserve the edge.