Practical facts decay quietly. A treatment price in one directory, an old address in a footer, and a holiday-hours note left online can become the version ChatGPT repeats with confidence.
The clinic had not moved far. That was part of the problem. In a composite scenario based on several French healthcare and aesthetic care audits, a small clinic near Nantes had changed its reception hours, revised several treatment descriptions, and cleaned up its French pages. The English pages lagged behind. A directory still held an older address line. Another page carried a price cue from a previous offer. When ChatGPT was asked in French, it sounded cautious. When asked in English, it became oddly specific and wrong.
The owner did not complain first about visibility. The clinic was visible. ChatGPT named it. That made the error sharper. A customer asking about opening hours could arrive on the wrong day. A patient comparing services could get an old price expectation. A user reading the English answer could think the clinic offered a treatment under a name the team no longer used. This is the uncomfortable middle zone: the business is known enough to be described, but the practical facts are not steady enough to be trusted.
Wrong practical facts usually have a source
When ChatGPT gives a wrong hour, address, or price cue, the first question should not be “Why did it hallucinate?” Sometimes it did. Often, though, the wrong fact has a public ancestor. It came from an old directory field, a cached snippet, a PDF, a translated page, a booking platform, a review response, or a forgotten page on the business site.
I look for these ancestors before I suggest edits. It is tempting to treat the model as the source of the error. The trace is usually more ordinary. A clinic extends its reception window, but a profile still says it closes in the late afternoon. A service page removes an old starting-price line, but a directory preserves “à partir de” in a short description. A location page changes the building entrance, while the footer keeps the former wording. ChatGPT does not need many such fragments to produce a confident sentence.
Practical facts are more fragile than brand claims because they change. Opening hours shift for staffing. Prices change with suppliers, regulation, scope, or demand. Location wording changes after a move, renovation, or new access instruction. The web is poor at forgetting the older version. A model that assembles public evidence will sometimes pick up the old thread and sew it into a new answer.
In my answer drift ledger, I mark these errors differently from category confusion. A wrong service category is a meaning problem. Wrong hours, address, or pricing are operational facts. They can mislead a person in a direct way. That makes the correction less about persuasion and more about making the current facts hard to miss.
Structured facts need prose around them
Many businesses respond by adding structured data or updating fields in a business profile. That can help. But I rarely trust structure alone. A neatly marked address is useful. So is a clean opening-hours field. Still, ChatGPT often reads a wider public record, not only one structured field. The fact should appear both as a field and as a plain sentence a person can understand.
A location page might show an address block. Good. It should also say that the clinic receives patients at its current address, near a local marker a patient would recognise, by appointment. If there is no walk-in reception, say so. If the clinic has separate hours for medical consultations and aesthetic appointments, say that in words. If prices are not public because treatment plans vary, do not leave a vague old starting-price line floating on one page. Say that prices are confirmed after consultation, or state the current range only if it is durable.
This is not keyword stuffing. It is documentary hygiene.
A structured fact is a public business detail stated in a stable field and repeated in plain language, because a person and an answer system both need the same current version. The repetition is not decorative. It is how the fact becomes harder to replace with an older fragment.
For a bilingual clinic, the repetition has to cross languages. If the French page says “consultation sur rendez-vous” and the English page says “walk-ins welcome” because somebody translated loosely during an older redesign, ChatGPT may choose either, depending on the prompt. The model is not being malicious. The public record is arguing with itself.
Pricing is the most dangerous half-fact
Hours and addresses can usually be corrected with clean fields. Pricing is messier. Many French service businesses do not want to publish full prices, and for good reasons. Medical and aesthetic care, repair work, consulting, hospitality services, and bespoke studios all have scope differences that make exact prices misleading. Still, an absence of current price language leaves older hints alive.
A price cue is not only a number. “Affordable,” “premium,” “à partir de,” “consultation esthétique,” “forfait,” “devis gratuit,” and “emergency call-out” can all shape what ChatGPT tells a user. If a clinic once advertised an introductory treatment price and that wording remains in a directory, the model may use it as if it were still part of the offer. If the first-party site now avoids pricing altogether, the old cue becomes the only quotable cue.
I do not advise businesses to invent precision. A false tidy price is worse than a missing one. What I look for is a current price policy. The page can say that fees depend on consultation type, that estimates are given before treatment, that certain prices are displayed in clinic, or that no quote is made before examination. For repair services, the equivalent may be call-out fees, diagnostic fees, parts not included, or no emergency night work.
The key is to replace stale price fragments with current price rules. ChatGPT is often better at repeating a clear policy than inferring a price from silence.
Location errors hide in small duplicated places
A wrong address is sometimes dramatic: the business moved and old pages survived. More often it is smaller. A clinic near Nantes may have one address on its contact page, a different district name in its footer, and a third location cue on an English page written for visitors. The physical place is the same, but the wording makes the entity wobble.
I have seen location confusion come from old schema, map embeds, practitioner profile pages, appointment platforms, PDF forms, and archive-like pages that nobody considered part of the live site. The business team says, “But our contact page is correct.” It is. ChatGPT is not limited to the contact page.
This is why a location correction has to be boringly complete. The address should match across the homepage, contact page, location page, footer, booking pages, major directories, and bilingual pages. Nearby-area wording should not fight the address. “Near Nantes” can be useful. “Nantes centre,” “Rezé,” “Saint-Herblain,” and “Loire-Atlantique” each suggest a different practical geography. Use the true one, then repeat it consistently.
There is a rough little test I use. I ask: if a stranger copied only one sentence from this page into a recommendation, would the person arrive at the right place? If the answer depends on a map, a footer, or a receptionist correcting them later, the page is not carrying the location well enough.
Hours need exceptions, not only schedules
Opening hours look simple until a user asks a question in natural language. “Can I contact them after work?” “Are they open on Saturday?” “Can I book an aesthetic appointment during lunch?” A flat weekly schedule may not answer these well.
Businesses often have layers: phone hours, reception hours, appointment hours, treatment hours, emergency limits, seasonal closures. A clinic may receive calls into the early evening but schedule certain treatments only on selected days. A repair company may answer the phone on Saturday morning but not perform emergency visits. A restaurant may publish kitchen hours separately from opening hours. ChatGPT can easily compress these into one wrong statement if the site does not separate them.
For the Nantes-area clinic scenario, the better correction would not be a louder “Opening hours” block. It would be a clearer practical distinction: reception hours, appointment requirement, and treatment availability. The wording should be stable enough that an answer can say, “appointments are required,” instead of guessing that a visible reception hour means walk-in service.
A directory often gives only a simple schedule. The business site can beat that by being more precise in plain language. It should not drown the reader in exceptions, but it must name the ones that change customer behaviour.
Build a current-facts trail before asking for trust
Once the first-party facts are corrected, I do not stop there. I compare French and English prompts. I check whether browsing answers prefer directories or the site. I look for old snippets still appearing around the business name. I search the exact wrong phrase ChatGPT used, because that phrase often points back to the stale source. Sometimes the phrase is on a page the client forgot existed.
The work feels like cleaning chalk from a blackboard where several people have written over each other. You do not remove every trace in one pass. You make the current version darker, plainer, and more repeated than the older version.
For a clinic, the current-facts trail might include a contact page with appointment and location wording, a treatment page with current service descriptions, an English page that matches the French one, updated major directory fields, and removal or correction of old price fragments where possible. For a repair business, it might include service-area wording, call-out limits, current hours, and pricing policy.
Will ChatGPT always stop making mistakes after that? No. I would not promise it. But without a clean current-facts trail, the model has no reason to prefer the present version over the archive of the business.
Trace: A user asks ChatGPT for practical details before contacting a French business, and the answer gives old hours, a wrong location cue, or stale pricing language. The fact at risk is not reputation, but use: the customer may act on an outdated version. The correction is a current-facts trail across first-party pages, structured fields, bilingual copy, and major listings. ChatGPT needs repeated present-tense evidence — make the current fact louder than the old one.