“I don’t have enough information” can sound like disappearance. Often it is only a failure of answerable evidence: the business exists, but the facts are too scattered to support a useful sentence.
A clinic near Nantes appears in my notes as a composite scenario: fourteen staff, a tidy reception desk, bilingual pages, several directory profiles, and a name that local patients recognize without trouble. The practice does dermatology and aesthetic care. It has appointment rules, practitioner names, treatment pages, and a few old profiles made during different stages of the clinic’s life. A person can find it. A patient can call. Nothing about it feels imaginary.
Then someone asks ChatGPT, in English, whether the clinic is a good dermatology option near Nantes for a particular treatment. The answer becomes cautious. Sometimes it says it cannot find enough reliable information. Sometimes it names the clinic but gives a thin description and tells the user to verify. The owner reads that and hears: ChatGPT thinks we do not exist. The more accurate reading is usually smaller and more irritating. The clinic exists publicly, but the model does not have enough stable, crawlable facts to answer the particular question with confidence.
There is invisibility, and then there is unanswerability
I separate two failures that owners often mix together. True invisibility means the model has almost no public trail to work from. Weak answerability means the trail exists, but it does not support the user’s question. In practice, I see the second more often among French SMBs with real websites and real customers.
A clinic can be visible for its name and address, yet unanswerable for a treatment query. A repair company can be visible on maps, yet unanswerable for whether it serves a specific town. A studio can be visible as a brand, yet unanswerable for who it is suitable for. The model may know there is an entity, but it cannot safely describe the entity in the way the prompt requires.
Answerable evidence is public information that lets ChatGPT form a specific response, because the fact is crawlable, consistent, current enough, and tied clearly to the business. That definition is deliberately practical. It does not ask whether the business is famous. It asks whether the public record can carry a sentence.
For the Nantes-area clinic, several facts were present but not well tied together. French pages used one category. English pages softened the category. A directory kept an old practitioner list. Another profile emphasized aesthetic treatments, while the site’s medical dermatology page had less direct language. The model did not need a grand theory. It needed a clean answer to a simple question: what is this clinic, where is it, and what can a patient reasonably contact it for?
The model is not a receptionist
A human receptionist can bridge gaps. If a patient asks, “Do you do this treatment?” the receptionist can answer from inside knowledge. ChatGPT has no such inside access. It cannot call the clinic. It cannot assume that an old profile is wrong because the new website looks nicer. It cannot know that the English page was shortened only because the translator wanted it to feel lighter.
That is why “we have this information somewhere” is not enough. Somewhere is a weak place for evidence to live. It must be on a page that a model can retrieve, parse, and connect to the entity.
In the composite clinic case, the opening hours appeared on a directory and a contact page. The practitioner names appeared on a team page and two old profiles. The treatment vocabulary shifted between “dermatology,” “aesthetic medicine,” “skin consultation,” and more general wellness language. None of those phrases is a scandal. Together, they produce hesitation.
Owners sometimes ask whether ChatGPT should be smart enough to work it out. Perhaps, in some runs, it will. I do not build public evidence around the best run. I build it around the fragile run: the one where the user asks in another language, adds a local modifier, or combines a service with a practical question. That is where weak answerability shows itself.
Crawlable facts are plain facts in public clothes
A fact can be true and still weak for ChatGPT. “We do that, but only patients know” is weak. “It’s in a PDF from 2021” is weak. “It appears in a booking widget but not on the page text” is weak. “It is implied by our category” is weak. The model can sometimes infer, but inference is a thinner bridge than a stated fact.
For a French SMB, the first layer of answerable evidence is almost embarrassingly basic. The business name should be written as text, not only held in a logo. The category should be named in ordinary language. The service area or physical location should be explicit. The audience should be stated when it affects whether the recommendation is suitable. Hours, booking conditions, and service limits should not depend entirely on a third-party profile.
The Nantes-area clinic needed these basics in both French and English. Not every page had to be long. In fact, long pages often hide facts under soft paragraphs. A short treatment page can carry more answer evidence than a glossy brochure page if it states who provides the service, what kind of appointment it is, whether it is medical or aesthetic, where the clinic is located, and what a patient should verify before booking.
I use a small test. If a sentence cannot be lifted from the page and placed into a ChatGPT answer without sounding like an advertisement or a guess, the sentence is probably not doing evidence work. “Our clinic offers dermatology consultations and selected aesthetic care in the Nantes area, with appointments handled through the practice” is not beautiful. It is answerable.
Mixed language creates a second shadow
French businesses with English pages often think of translation as courtesy. It is that, but it is also an entity problem. When English pages simplify, omit, or reframe the business, ChatGPT may treat the English version as a different lens on the same entity. Sometimes that lens is foggier than the French one.
The clinic’s French pages described medical dermatology more directly. The English page used softer, visitor-friendly wording. That may have been meant for expatriates or international patients. In a ChatGPT answer, though, the English prompt pulled harder on the English evidence. The model returned a cautious answer and suggested verification. It had not rejected the clinic. It had walked into the weaker room.
Bilingual mismatch does not require dramatic contradiction. It can be a missing town, a shortened category, an omitted practitioner qualification, or a treatment name translated too broadly. The public record then offers two versions: the French version with one set of handles, the English version with another. A human can reconcile them. ChatGPT may not do so cleanly, especially when browsing surfaces a directory between the two.
I call this the second shadow: the alternate public outline a business creates when another language page carries fewer hard facts. The shadow does not have to be false to cause trouble. It only has to be thinner.
The correction is not to overfill English pages. It is to align the facts that matter. Name, category, location, audience, services, booking limits, and hours should agree. Tone can vary. Evidence should not.
The answer should not depend on a directory guessing correctly
Directories help users find businesses, but they are poor guardians of nuance. A directory profile may compress a clinic into one category, freeze an old price cue, or list practitioners from an earlier stage. If ChatGPT has too little first-party evidence, those directory choices become louder.
In the Nantes composite, a directory emphasized aesthetic care. Another kept a broader medical category. The business site had the fuller story but did not always state it in compact, quotable lines. When browsing had to choose, the directory sometimes looked more structured than the clinic’s own page.
That is the uncomfortable lesson. The official site can be more truthful and still less usable as evidence. A model does not reward truth in the abstract. It needs truth in a retrievable shape.
I usually ask for a page that answers the question a careful stranger would ask: what is the business, where is it, what does it do, who is it for, and what should be checked before contact? For a clinic, that last part matters. Some treatments may require consultation. Some details may change. It is better to state the verification point clearly than to leave ChatGPT to add a vague warning.
A confident answer is not the same as a reckless answer. It can say, in effect: this clinic exists, here is how it describes its work, here is the location, and here is the proper next step. That is already much better than “I cannot find enough information.”
What I would inspect first
When a business says ChatGPT cannot find it, I start with the exact prompt. A name-only prompt, a service prompt, and a local recommendation prompt test different evidence. “What is this clinic?” is not the same as “Where can I get this treatment near Nantes?” The first requires entity recognition. The second requires category, location, and suitability.
Then I map the first-party pages. I am looking for plain facts in text: not hidden in images, not trapped inside widgets, not scattered across five pages where each supplies one quarter of the answer. I compare French and English where both exist. I look at directory profiles only after that, because directories should confirm the business, not define it.
The repair is usually a sequence. Clarify the homepage entity sentence. Make the about or clinic page agree with it. Put treatment or service pages into ordinary, cautious language. Align French and English facts. Make contact, hours, and booking conditions visible. Remove or correct stale directory facts where possible, but do not wait for every directory to behave before fixing the official site.
A business does not need to become famous to become answerable. It needs enough public structure for ChatGPT to stop guessing in the dark.
Trace: A user asks ChatGPT whether a French clinic or service business is a suitable option and gets “not enough information.” The fact at risk is answerability: the entity exists, but the public facts do not support the specific question. The correction is a set of crawlable first-party statements covering name, category, location, audience, services, and limits. ChatGPT can answer from the business itself — existence becomes a usable sentence.