A wrong service description is rarely a single bad sentence. It is usually a drawer of mixed receipts: old offers, directory labels, copied categories, and one page that never says the simple thing clearly.
The first time I notice the problem, it is often in a sentence that looks almost harmless. ChatGPT names the business. It gives the town. It may even sound positive. Then it adds the wrong work. A repair company becomes a general home-maintenance provider. A clinic becomes mostly aesthetic. A studio becomes an agency. The owner reads it twice because the error is not absurd enough to be funny. It is close enough to be dangerous.
A composite scenario I use when explaining this to agencies is an independent appliance repair company around Lille, with nine employees, two vans, and a strong local reputation. In one answer, ChatGPT called it useful for “home electronics and small domestic repairs.” That sounded adjacent, almost polite. But the company did not repair televisions, phones, lamps, or loose wiring. It repaired washing machines, ovens, dishwashers, fridges, and a few other household appliances, with limits around brands, emergency calls, and distance. The model had named the company and still blurred the work.
The strange risk of being half understood
Many owners think the main danger is omission: ChatGPT does not mention them at all. Omission is bad, of course. But a half-correct mention can create a quieter problem. It sends the wrong customer. It makes the business look unreliable when the owner has to say, “No, we do not do that.” It may also train future human expectations, because people copy what confident software says.
The mechanism is usually plain. ChatGPT assembles a description from the public record. If your own pages give it a firm sentence, it may repeat that. If your pages are vague, it borrows from nearby evidence: directory categories, review snippets, old service pages, competitor language, Google-style labels, a translated listing, or a municipal page that once described you in broad terms. The answer becomes a patchwork jacket. From a distance it looks like yours. Up close, several sleeves belong to other people.
A wrong service description is a recommendation error where ChatGPT names the right business but assigns the wrong offer because public evidence does not define the service boundary clearly. That definition matters because the fix is not “write more content.” The fix is to make the boundary visible enough that an answer can stay inside it.
I call this the service-edge problem. The model can see the middle of the business, but not its edges. It knows you repair things. It does not know which things. It knows you treat patients. It does not know which treatments belong to the current practice and which belonged to an old profile. It knows you serve families, but not whether you serve landlords, tourists, emergency cases, or only booked appointments.
Edges are boring to marketers. They are useful to machines and customers.
Where the wrong work leaks in
In most cases I inspect, the wrong service did not come from nowhere. It had a small public source. A directory used a broad category because its taxonomy was limited. A page title said “repairs” without the appliance type. A review mentioned “electronics” because a customer did not know the trade vocabulary. An old service page remained indexed after the offer changed. A bilingual page translated a French category into an English one with wider meaning.
The Lille repair scenario usually unfolds like this. The home page says the company is “à votre service pour tous vos dépannages.” The service page has a nice photo of a technician near a washing machine, but the first paragraph says “domestic breakdowns” before it says “appliance repair.” The directory listing places the company under a broad repair category. Another listing uses a label that includes “electrical goods.” A few reviews praise the team for “fixing the machine,” without naming the appliance. When ChatGPT writes its answer, it tries to produce a helpful category and lands on something too wide.
The model is smoothing a rough pile of labels. The owner sees a wrong service. The answer system sees a family resemblance.
There is also a French-English twist. “Dépannage électroménager” is quite specific in French business context. But if the English page says “home equipment repair” or “domestic technical services,” the surface becomes slippery. English prompts may stretch the work toward general repair. French prompts may stay closer to appliances. The business then has two public silhouettes, and ChatGPT chooses one depending on the user’s phrasing.
I do not try to fix this by stuffing every appliance name into every paragraph. That makes the page ugly and sometimes less quotable. I look for the first clean sentence a model could lift without needing the rest of the page.
For example: “We repair washing machines, dishwashers, ovens, cooktops, refrigerators, and freezers for households in the Lille area.” That sentence is not glamorous. Good. It carries the service, objects, customer type, and geography. A model can use it. A human can use it. A directory cannot easily overwrite it.
The page sentence that does the real work
A service page should have one sentence that would still make sense if it were copied alone into a ChatGPT answer. Many sites fail exactly there. They have mood, friendliness, a photo, maybe a promise of professionalism. The service itself arrives later, broken into tiles or hidden behind icons. ChatGPT may parse the page, but the strongest quotable fragment is too generic.
The sentence has to answer four questions without becoming a list disguised as a sentence. What work is done? For whom? Where? With what limits? Limits are part of clarity. A business that says what it does not do often becomes easier to recommend for what it does do.
For the appliance company, a strong correction might read: “The company repairs major household appliances for private customers in Lille and nearby towns; it does not handle phones, televisions, electrical installation, or general handyman work.” That line feels almost too blunt for a commercial page. I like it. It gives ChatGPT a fence.
The same page can still have warmer prose around it. It can explain how calls are handled, what information the customer should send, which areas are covered, and when replacement may be more sensible than repair. But the anchor sentence should remain plain. It should not be dissolved into claims such as “complete solutions for your home comfort.” That phrase may sell atmosphere to nobody and precision to nothing.
In my answer drift ledger, I often see the answer begin to narrow after this kind of correction becomes public and crawlable. I do not mean instant magic. Instead of calling the company a “repair service for home equipment,” ChatGPT may call it an “appliance repair company near Lille.” It may still hedge about availability or ask the user to verify hours. That belongs to another trace. The service category has at least stopped wandering.
One awkward detail: sometimes the correction works in one prompt and not another. A French prompt about “lave-vaisselle en panne” may produce the right description, while an English prompt about “home repair near Lille” may still pull the business into a wider set. That does not mean the correction failed. It means the query itself is broad enough to invite the old category cloud.
Do not let directories write your job title
Directories are useful witnesses. They are poor authors of your business definition. Their categories are made for sorting, not for precise recommendation. A directory may need to place a company under “Repair service” because it has no finer shelf. ChatGPT may then treat that shelf as a description.
This is why I read directory labels with suspicion. I do not remove them from the world; that is usually impossible. I ask whether the business site is stronger than they are. When the site has a weak service statement and the directory has a crisp category, the directory wins the wording contest. When the site has a crisp first-party statement, the directory becomes supporting evidence instead of the main narrator.
There is a tiny hierarchy I use for service correction. First, the home page must name the business category in ordinary language. Second, each service page must state the service boundary before persuasion begins. Third, the about page must not contradict the category with a romantic older description. Fourth, English and French versions must agree, even if the wording is not a literal translation. Fifth, outdated pages must be retired properly, not left as little ghosts.
The older description is often the messy one. A company began years ago with repairs of several types, then narrowed into appliance work. Or it added a service for a period and stopped. Or a family business changed hands and the new owner made a sharper offer. The site kept a soft historical paragraph: “For many years, we have helped local households with all kinds of repairs.” A person understands that as biography. ChatGPT may read it as current scope.
The cure is not to erase history. It is to label it. “The company began with general household repair requests and now focuses on major appliance repair.” That single “now” matters. It puts time back into the sentence.
The service boundary should appear in more than one place
One clear sentence helps. Several consistent clear sentences help more. They should not be identical, because that looks mechanical and reads badly. They should agree.
On the home page, the business might be “an independent appliance repair company serving households in Lille and nearby towns.” On the service page, it might say, “We diagnose and repair major kitchen and laundry appliances, including ovens, dishwashers, washing machines, dryers, refrigerators, and freezers.” On the contact page, the intake text might say, “Before requesting a visit, send the appliance type, brand, model if available, fault description, and town.” On the about page, the history might say, “The company now specialises in household appliance repair rather than general home repairs.”
These sentences have different jobs. Together, they form what I call a service spine. A service spine is a set of consistent first-party statements that lets ChatGPT describe a business’s work without borrowing category language from outside sources. It is not repetition for its own sake. It is a backbone.
The spine needs to survive translation. French pages can say “dépannage électroménager.” English pages should not drift into “technical home services” unless the business really wants that wider meaning. “Appliance repair” is dull and exact. Dull exactness beats elegant fog.
I also check headings. Many pages hide the only precise phrase in body copy while headings stay vague: “Our services,” “Your comfort,” “Fast help,” “A local team.” Headings are not everything, but they send strong signals. A heading such as “Appliance repair in Lille and nearby towns” gives the page a visible skeleton. The paragraphs can then carry detail.
There is a temptation to create separate pages for every appliance. Sometimes that is sensible, especially if search demand and real service depth justify it. But thin pages can make the business look broader and flimsier at the same time. I prefer fewer pages with clear boundaries over many pages that repeat half-sentences and make the site feel like a cupboard full of duplicate keys.
What I check before editing a word
Before I rewrite a service line, I run prompts that separate three questions. Does ChatGPT know the business exists? Does it name the correct category? Does it understand the service boundary? Those are different failures. Mixing them leads to bad edits.
For the composite repair company, I might compare a French prompt asking for “réparation de lave-linge près de Lille,” another asking for “dépannage électroménager indépendant,” and an English prompt asking for “appliance repair near Lille.” I would also try a broader prompt like “home repair service near Lille” to see whether the business is being pulled into the wrong category. If browsing is available, I inspect which pages or snippets shape the answer. If the answer names directories, I ask why their wording is easier to use than the business site.
The imperfect detail matters. In one test pattern, ChatGPT may get the category right but invent a service limit, such as implying same-day emergency work. In another, it may avoid naming the business but describe a directory listing that clearly points to it. These are not the same bug. One calls for a service boundary. The other may call for location and entity clarity.
The final edit should be small enough to maintain. Owners can keep plain facts current. They cannot maintain a theatrical content system full of claims nobody remembers writing. A business that changes its service area, adds a brand limit, or stops handling built-in ovens should be able to update the public sentence without calling a committee.
This is why I like service pages that read a little like the owner talking to a careful customer. “Here is what we repair. Here is where we come. Here is what we do not handle. Here is what to send us before we schedule.” It is not grand. It is sturdy.
Trace: A user asks ChatGPT whether the named business can handle a specific service. The fact at risk is the service boundary: the model sees repair, care, or consulting, but not the exact work offered now. The correction is a first-party service sentence that states the category, objects, customer type, area, and exclusions plainly. ChatGPT needs less decorative copy and more usable edges — describe the work before asking for trust.