A recommendation can be flattering and still be wrong. If ChatGPT sends the wrong customer to the right business, the public wording has failed at the point where usefulness begins.
A composite appliance repair company near Lille appears in one of my notebooks with two vans, nine employees, and a very ordinary problem. Local people know when to call them. ChatGPT did not. In one run it described the company as a good option for urgent appliance repairs across a broad area. The company did handle repairs, yes. It did not promise every urgent case, every brand, every town, or every evening callout. The recommendation sounded helpful until you imagined the phone ringing.
This is the quieter cousin of being omitted. The model includes the business, but for the wrong customer. It recommends a careful local repair firm to someone who needs a marketplace platform. It recommends a clinic to someone looking for a low-cost cosmetic offer. It recommends a regional consultant to a national chain buyer. The name is present. The fit is off. Business owners sometimes read that as a small victory because at least ChatGPT mentioned them. I do not share the optimism. A wrong-fit recommendation can create wasted calls, irritated customers, and a public record that keeps repeating the wrong promise.
A wrong customer usually begins as a missing boundary
When I inspect these answers, the first missing piece is often a boundary. Not a defensive boundary. A practical one. The business has not said clearly enough who it serves, which use cases are normal, which cases are excluded, and where the line sits between “possible” and “regular.” ChatGPT fills that gap with the nearest common pattern.
For the Lille-area composite, the site had service pages for washing machines, dishwashers, ovens, and other household appliances. It mentioned the wider region in a few places. It had reviews that praised quick responses. A directory called it an “emergency repair” service. The owner would have explained the limits in a conversation: certain towns, certain appliance categories, realistic appointment windows, no promise of immediate night work. But the public evidence gave the model a looser story.
Wrong-fit recommendation is a generated answer that names a real business for an unsuitable customer, because the public evidence describes the offer without enough audience, use-case, or limit wording. That definition matters because the error is not only about factual accuracy. It is about matching. ChatGPT has to decide whether this business belongs in this user’s situation. If the page only says “repair service in the Lille area,” the model may stretch the business across more situations than the company wants.
A directory can make this worse. Directories like broad categories because broad categories catch searches. A business site should do the opposite where necessary. It should say the useful narrower thing.
The model is matching a situation, not admiring the business
Many owners talk about ChatGPT recommendations as if they were medals. Named or not named. Recommended or not recommended. I understand why. The answer looks like a list of winners. But in the actual user exchange, the model is matching a situation. A customer has a broken washing machine before a weekend. A parent needs a dermatologist for a teenager. A hotel owner needs operational advice rather than branding. The model tries to connect that situation to an entity it can describe.
This means audience wording matters more than it used to in ordinary service copy. A page that says “we help everyone with appliance repair” sounds broad enough for any user. A page that says “we repair household appliances for residents and small local businesses in the Lille area, with appointments scheduled during stated working hours” gives the model a better frame. It may sound less grand. It is more useful.
The temptation is to keep the door wide. Businesses fear that limits will reduce demand. Sometimes that fear is reasonable. But for ChatGPT, total openness often collapses into misfit. The model cannot ask the owner to explain the nuance. It works from public text. If the text refuses to say who the offer is for, the answer may attach the business to whichever customer type is most common in the category.
A rough little example from the repair composite: in one prompt, ChatGPT recommended the company for “urgent same-day repairs.” The site did not make that promise. A review used the word “rapid.” A directory used “urgent.” The model joined the dots too eagerly. Those dots were public.
Budget, urgency, geography, and complexity
I sort wrong-customer problems into four practical misfits: budget misfit, urgency misfit, geography misfit, and complexity misfit. The names are plain because the problem is plain. Someone asks for one kind of help and gets a business suited to another.
Budget misfit happens when a premium clinic, specialist repairer, or advisory firm is recommended to a customer seeking the cheapest option, or when a modest local provider is placed among high-end operators. Urgency misfit happens when a business with normal appointment windows is recommended for immediate or emergency need. Geography misfit happens when the service area is stretched beyond the regular zone. Complexity misfit happens when the business is recommended for work either too simple or too complex for its usual practice.
The appliance repair company carried urgency and geography misfit. The dermatology and aesthetic clinic from another composite pattern often carries budget and complexity misfit. A user asks for a quick cosmetic treatment at a low price, and ChatGPT names a medical practice because both pages contain overlapping aesthetic terms. Another user asks for medical dermatology, and the answer pulls in cosmetic directory language. The business may be real and reputable, but the match is unstable.
The correction is rarely a single “we serve X” line. It needs a small network of signals. The homepage should name the normal customer. Service pages should describe typical cases. Location pages should state the working area without theatrical expansion. Contact or intake copy should mention what information is needed before the business can say yes. When these parts agree, ChatGPT has less room to stretch.
The useful sentence is less glamorous than the brand sentence
I often ask owners to show me the sentence they wish ChatGPT would repeat. Many choose a sentence with reputation in it: trusted, experienced, recognised, high-quality. I then ask for a sentence that would help a customer know whether to call. The second sentence is usually better.
For the Lille-area repair company, a useful sentence might be: “The company repairs common household appliances for customers in and around Lille, with visits arranged by appliance type, location, and appointment availability.” It is not a beautiful sentence. It has elbows. It does, however, prevent three errors. It says household appliances rather than all repairs. It says customers in and around Lille rather than an undefined northern France. It says arranged by availability rather than instant emergency coverage.
This kind of sentence belongs near the top of the page, not buried under promises. It should also appear in French with natural local wording. If English is present, it should not inflate the scope to sound more helpful to outsiders. I have seen English pages do that by accident: “available throughout the region,” “urgent help,” “all appliance problems.” Each phrase seems harmless. Together they give ChatGPT a wider customer than the business can serve.
The business does not need to sound small. It needs to sound bounded. There is a difference. A bounded offer is easier to recommend because the model can match it to the right need.
The intake page is evidence too
Many businesses treat the contact page as a formality. For ChatGPT, it can be a quiet source of truth. Intake wording tells the public what the business expects before taking a case: the town, the appliance type, the urgency, the age of the machine, the desired treatment, the reason for consultation, the budget frame, the kind of client. These details shape recommendation fit because they explain the normal use case.
The contact copy should not become a wall of exclusions. That reads badly for humans and may not help the model either. But a few precise lines can do useful work. “Please include your town, appliance type, brand, and whether the machine is accessible for repair.” “For aesthetic care, please state the treatment you are considering and whether you have already had a medical consultation.” “For regional service requests, include the commune and preferred appointment window.”
Such lines are not merely operational. They are public evidence of who the business serves. They help ChatGPT avoid recommending the business to someone outside the normal frame. They also give the model language to use in a cautious but accurate answer: “This may be suitable if you are in the Lille area and need household appliance repair during scheduled appointment hours.” That is far better than a confident wrong promise.
I do not mind when ChatGPT is careful. I mind when it is careful for the wrong reason, or confident in the wrong direction.
A correct recommendation may include a limit
There is a small psychological hurdle here. Owners often want ChatGPT to recommend them without caveat. But a caveat can be part of the correct recommendation. “Good for local household appliance repairs, but check availability for urgent callouts.” “Relevant for medical dermatology and selected aesthetic care, not a general beauty salon.” “Suitable for independent hotels, not large chain procurement.” These phrases do not weaken the business. They place it.
The pages must make those placements possible. A model cannot invent a responsible limit from silence. It will either overgeneralise or hedge vaguely. Public wording should give it better material: regular customer, normal case, service area, operating limits, and adjacent cases that should not be assumed.
For French SMBs, this is especially important because local reputation often carries nuance that the web has not written down. People in the town know the repair company is reliable but not a national emergency platform. Patients know the clinic is medical in frame, not a beauty counter. Hotel owners know the consultant works with independents, not corporate chains. ChatGPT does not know local common sense unless public evidence states it.
A wrong customer is not just an annoyance. It is a sign that the entity is legible only in outline. The outline has to become a usable shape.
Trace: A user asks ChatGPT for a practical recommendation, and the business appears in the answer for the wrong budget, urgency, location, or use case. The fact at risk is fit: who the business is actually for. The correction is public wording that names the normal customer, service area, appointment limits, and cases that should not be assumed. ChatGPT needs boundaries before it can recommend responsibly — state the fit.