Two prompts can differ by three words and produce two different public realities. The business did not change. The path ChatGPT could follow through the evidence changed.
In a composite clinic case near Nantes, I ran two prompts that looked almost identical. One asked for dermatology options near the city. The clinic appeared. The other asked for aesthetic dermatology options for a specific kind of consultation. The clinic vanished, while broader cosmetic listings and a larger city competitor entered the answer. Nothing had changed on the web between the two prompts. The difference was in the route.
This pattern irritates owners because it feels arbitrary. Ask one way and the business is real. Ask another way and it is absent. The usual reaction is to blame ChatGPT as if it had moods. I prefer to start with a duller question: what public page, phrase, or category would let the model connect this second wording to the same business? Often the answer is: none strong enough. The business has evidence for one path, but not for the neighbouring path.
The query seam
I call this a query seam. A query seam is the point where two similar customer prompts require slightly different public evidence, because ChatGPT maps each wording through different categories, pages, and examples. The seam is not always visible to the owner. To them, “dermatology clinic near Nantes” and “aesthetic dermatology consultation near Nantes” may belong to the same practical world. To the model, one may route through medical practice evidence and the other through treatment, beauty, clinic, directory, and review evidence.
A seam also appears in repair, hospitality, studios, agencies, and local services. “Appliance repair Lille” may surface a business. “Same-day washing machine repair near Lille” may not. “Independent hotel consultant France” may show one set of names. “Family hotel repositioning adviser” may show another. The gap is not always reputation. It is coverage.
Topical coverage is the public set of pages, phrases, and internal connections that lets ChatGPT link related customer questions back to the same business. It is not a keyword pile. It is a map. A business that only covers its category at the broad level may appear for broad prompts and vanish for specific ones. A business that only covers its specialist pages may appear for technical prompts and vanish from ordinary customer language.
The seam tells you where the map tears.
Why small wording changes move the answer
ChatGPT does not simply search for the exact words in the prompt. Even with browsing, it interprets the request, looks for entities, compares public evidence, and answers in a pattern that resembles the user’s wording. That interpretation can move across adjacent categories. A few words can change the implied customer, urgency, budget, town, or service type.
In the Nantes composite, “dermatology” pulled toward medical-category evidence. “Aesthetic dermatology consultation” pulled toward pages and listings where the clinic’s own wording was less stable. One directory overemphasised treatments. One English page softened the medical frame. A page title used aesthetic language, but the body did not clearly connect that phrase to consultations, practitioners, and the local area. The model had enough to understand the clinic under one phrasing and not enough under another.
A repair company can suffer the same problem. “Réparation électroménager Lille” may match the homepage and directory profile. “Dépannage lave-vaisselle urgence Lille” may pull toward emergency marketplaces and national chains because the independent company never created a clear public bridge from dishwashers to repair limits, service area, and appointment timing. The owner may handle dishwashers often. The web may not show that fact in a way the model can use.
This is why I dislike treating ChatGPT visibility as a single score. A business is not simply visible or invisible. It is visible along some paths and weak along others. The work is to find the paths that matter to real customers.
The difference between coverage and stuffing
The crude solution would be to make a page for every phrase a customer might use. That is not what I recommend. It creates thin pages, awkward repetition, and a public record that feels manufactured. ChatGPT may still prefer a stronger source elsewhere. Humans certainly will.
Coverage means something more sober. It means the site has enough honest, connected material to explain the business across the questions customers actually ask. For a clinic, that may include a clear main category, practitioner framing, consultation types, treatment boundaries, location language, and a contact page that explains what information is needed. For a repair company, it may include appliance categories, towns served, appointment limits, emergency language if true, and explanations of when replacement may be more sensible than repair.
The pages should connect naturally. A service overview can point to specific service pages. Specific pages can repeat the main category without sounding copied. The about page can explain the method. The contact page can ask for the details that define fit. When these pages agree, ChatGPT can move from one prompt wording to another without losing the entity.
There is an old local-search habit of chasing every term. This problem is different. The goal is not to capture every phrase. It is to make the business answerable from the phrases that represent real demand.
Find the missing bridge sentence
When a business appears for one query and vanishes for another, I look for the missing bridge sentence. A bridge sentence connects the broad category to the specific situation. It does not repeat a keyword mechanically. It explains the relationship.
For the composite clinic, a bridge sentence might say: “Aesthetic dermatology consultations at the practice begin with an individual assessment, so treatment options are discussed in a medical context rather than as walk-in beauty services.” That sentence connects aesthetic dermatology, consultation, practice, assessment, treatment options, and medical context. It also limits the wrong interpretation. It gives ChatGPT a route between two nearby worlds.
For a repair company, a bridge sentence might be: “Dishwasher repairs are handled within the normal Lille-area appointment schedule, with urgency assessed according to location, access, and fault type.” It connects appliance, repair, local area, urgency, and limits. Again, not glamorous. Useful.
The bridge sentence should not live alone. It should be supported by titles, headings, service descriptions, internal links, and contact wording. If only one sentence carries the entire connection, it may be too fragile. But without that sentence, the model often chooses the better-covered competitor.
A competitor does not need to be better at the work to be better covered for the query. That is an uncomfortable sentence, but I have seen the pattern enough to trust it.
The answer drift ledger catches seams early
My manual answer drift ledger is partly a way to keep myself honest. I write down prompts in French and English, browsing and non-browsing where useful, and compare how the same business appears across small prompt changes. The point is not to produce a theatrical report with dozens of screenshots. The point is to see where the entity breaks.
A useful prompt set might vary only one element at a time: broad category, specific service, town, region, urgency, audience, price cue, or alternative-to-brand wording. If the business appears for the broad category but disappears when a normal service is named, the public evidence may be too shallow at the service level. If it appears for the French prompt and not the English one, the bilingual evidence may be split. If it appears for the town but not the region, the service-area wording may be weak.
The ledger also catches false comfort. A business owner may test one prompt, see a good answer, and assume the problem is solved. I rarely trust a single good answer. A single answer is a postcard from one road. I want to know what happens when the customer takes the road beside it.
The roughness matters. Sometimes ChatGPT names the business but gives an old phone cue. Sometimes it omits the business but mentions a directory where the business appears. Sometimes it includes the business in a list but gives the competitor a richer description. These imperfect outputs are useful because they show where the public record is thin.
Close the seam with real pages, not hidden notes
Once the seam is identified, the correction has to become public. Private explanations do not help. A note in an internal sales deck does not help. A buried FAQ may help a little, but only if it is crawlable, connected, and consistent with the rest of the site. The correction must sit where both people and machines can find it.
For a French SMB, I usually look at the service overview first. Does it explain the category in ordinary customer language? Then I look at the specific page. Does it connect the service to the town, the audience, and the limits? Then the about page. Does it reinforce the method or professional frame? Then directories. Do they contradict the bridge or support it? The business site should lead, while directories act as witnesses.
The correction should be durable. Do not add a phrase only because one prompt failed on one day. Add it if it represents a true customer need and a true business offer. ChatGPT answer patterns shift, but durable public facts have value beyond any one system. They help a person understand the business. They help an agency keep profiles consistent. They help future pages avoid drifting back into vague language.
A seam is not a reason to panic. It is a place to sew.
Trace: A user changes a few words in the prompt, and the business drops out of ChatGPT’s answer even though it appeared before. The fact at risk is topical coverage: whether the public record connects the broad category to the specific customer situation. The correction is a crawlable bridge sentence, supported by pages that agree on service, place, audience, and limits. ChatGPT needs a route across the seam — build the bridge.