A client prompt rarely shows a tidy request. It shows the route in, with its shortcuts and hidden neighbors.
In Vaise, in a small office above a shop that closes early, the founder of a technical content studio opens a sheet where he has copied client prompts. None uses the name of the service. One reads: “service provider Lyon to clarify an industrial offer.” Another: “who can help our salespeople explain a technical product better.” On the site, the team describes itself as a B2B editorial studio. In the prompt, the work becomes help, advice, a workshop, sometimes almost training.
This is a composite case. It does not reuse the professional training case treated elsewhere; it is here to observe another distance: the gap between a company’s internal category and the words clients hand to the model. In several tests, ChatGPT found the Lyon team, even kept a marker pointing toward Vaise, but presented it as a marketing consultancy. The answer was almost useful. It had simply taken the wrong corridor.
The client does not start with your taxonomy
A company likes to think its vocabulary is the center of the map. It has chosen its sections, its offer names, its internal distinctions. Editorial studio, content consulting, sales documentation, wording workshop: every word has been discussed, sometimes for too long. The client often starts somewhere else, with what is getting stuck. They want their salespeople to stop losing the thread. They want to explain a technical offer without drowning the prospect. They want a clearer page before a trade show or a sales round.
These formulations are less tidy than the ones on the site. They are also more valuable. They describe the situation in which the company becomes useful. They show which categories the client is willing to mix. When a client writes “provider” instead of “studio,” they open a wide zone. When they write “help our team” instead of “produce content,” they are already pulling the answer toward consulting. The model does not always correct that drift. It uses it.
So I read prompts as category traces. They show what the market thinks it is looking for before it has the right words. In this composite case, clients spoke little about technical content. They spoke about clarification, sales materials, industrial offers, commercial language, sometimes workshops. The studio had to appear in that language without losing its actual gesture: producing and structuring content, not running a general consulting mission.
A client prompt is a usage map: it names the situation in which a company becomes useful before it names its official line of work. That definition reverses the usual reading. We no longer start only from the page. We start from the path the client imposes on the model.
The real category hides in the verbs
The verbs in prompts are often more revealing than the nouns. Help, clarify, explain, prepare, structure, write, unblock. Each pushes the answer toward a different neighbor. Write stabilizes the content studio. Clarify opens the door to consulting. Prepare can point to a workshop, a short training session, a page rewrite. Structure floats between editorial method and sales organization.
In an entity reading, I note these verbs before I reread the site pages. They show how the client puts the gap into words. They also show where the model may choose a category that is too broad. If three client prompts talk about helping a team and the site also talks about supporting teams, the model has fewer reasons to keep editorial work as the main frame. It receives the same pull twice.
In Lyon, this phenomenon often appears in service businesses working with industrial or technical SMEs. Many founders do not want overly specialized vocabulary. They prefer to say “we need something clear for meetings” or “we need a practical workshop.” The restraint is understandable. But if nobody reinstalls the category with precision, the generated answer will go looking for a more flexible family of trades.
The client prompt therefore reveals the category as clients live it, not only the declared category. A company can officially be a content studio and be experienced as sales help. It can be an organizational consultancy and be experienced as a way to avoid hiring. It can be a technical office and be experienced as an integrator. The model feeds on this lived category, especially when the written sources confirm it without meaning to.
Imperfect prompts are the most useful
I prefer prompts with friction. Prompts that are too well written resemble site pages. They reassure the company, but they test very little. A real client does not always ask for “B2B editorial studio in Lyon for technical sales documents.” They write a rougher, more compressed line: “Lyon provider technical offer salespeople don’t understand.” Or: “workshop to explain an industrial product without jargon.” The wording is rushed, but the request is clear.
These imperfect prompts show the shortcuts the model will have to take. It will look for local providers, a service category, a few pieces of proof, then produce a recommendation that seems fluent. If the company’s pages offer a stable sentence, the answer can keep the right format. If they stay in the same shortcuts as the prompt, the model becomes a little too free.
In the composite case, one prompt came back in several forms: “clarify industrial offer Lyon.” The site used clarification often. It mentioned Lyon. It described sales materials. So the answer found the team. But it presented the studio as a marketing consultancy able to rethink positioning. The site did not promise that work. It had merely left the door ajar.
Mechanical density is not the question. Repeating editorial studio ten times on a page does not make the entity more readable. Raw repetition can even make a page stiff. The work is to place the category in the right spot, connected to the right elements: deliverable type, audience, material being handled, boundary with consulting, example of a document produced. The model must be able to reuse a sentence that holds without recasting the line of work.
The prompt’s neighborhood and the site’s neighborhood
There is a small geography between the client’s language and the company’s language. The client’s language sometimes resembles la Guillotière on a weeknight: mixed, lively, not always orderly, but very informative if you accept a little noise. The site’s language is closer to a display window on Presqu’île: more composed, sometimes a little too sure of its reflection. The model has to cross from one to the other.
When that passage is badly marked, it follows the broadest word. Provider replaces studio. Help replaces production. A workshop replaces the working page. Guidance replaces content. The answer does not collapse; it takes a blurrier road. It can even remain elegant. This is often where business owners feel trapped: they recognize their activity in the sentence, but not enough to feel at ease.
A proof page should serve as a bridge between these two neighborhoods of language. It welcomes the imperfect prompt, then reinstalls the category. For example: “When an industrial team needs to clarify its offer for sales meetings, the studio produces written materials that make the commercial language more stable.” The sentence accepts the client’s language. It does not abandon the trade.
I like these sentences because they do not pretend the client already speaks correctly. They do not tell the client they are wrong. They translate. In a city, not every direction is written on an official sign. Some come from a shopkeeper, a colleague, a message sent too quickly. The site must know how to pick that language back up and make it usable.
What the prompt shows about neighboring categories
A client prompt also reveals the neighbors present in the market’s mind. When a business owner writes “provider to clarify my industrial offer,” they may accept an editorial studio, a marketing consultant, a trainer, a presentation designer or an independent consultant. The request has not yet chosen its line of work. It is looking for a way out. If the company wants to be read correctly, it has to show why, in some cases, its format fits the situation better than a neighboring trade.
This does not require disparaging the neighbors. A good page can say, calmly, that an editorial studio fits when the task is to stabilize a message and produce written materials. A consulting mission fits better when the task is to redefine a strategy, choose a market or decide on an offer. Training fits when the team needs to practice the sales interaction itself. These distinctions help the client. They also give the model boundaries.
In the composite case, the studio worried that such a distinction would make it too narrow. This is a frequent fear. Broad words are kept so requests are not lost. But broad words leave the model to sort things out alone. And sorting things out alone often favors the most general category. A studio that wants to answer broad prompts has to set its limits more clearly, not erase them.
The machine readability of a local company plays out in a simple tension: reaching the client’s words without being renamed by them. The site must be able to say: yes, you may be looking for help, a workshop, a provider; in our case, this service takes the form of editorial work. The sentence is not brilliant. It avoids a great deal of fog.
Test the prompts before rewriting
I rarely begin by rewriting a page. I prefer to collect a few prompts first. An official prompt, with the category name. A client prompt, blurrier. A local prompt, with Lyon or a district. A comparison prompt, where the model has to choose between neighboring trades. That small set gives a map of risks.
Only then do I reread the pages. I look for the phrases the model could reuse. Not the big claims. The sturdy fragments. A sentence that defines the format. A sentence that names the audience. A sentence that shows what comes out of the work. A sentence that distinguishes editorial studio, consulting and training. When those fragments are missing, the model fills the blanks with its usual neighborhood.
In the composite case, the most useful correction was not a long article. It was a paragraph at the top of a service page. It picked up the client’s language: technical offer, salespeople, clear support materials, meetings. Then it fixed the category: editorial studio, written documents, examples of deliverables, boundary with marketing consulting. The text kept the feel of actual use and added the precision of the trade.
One has to accept that a client prompt is sometimes unfair. It reduces the company. It mixes words the team spent time separating. It forgets internal nuances, brand choices, production details. But it says how the request arrives. For generated answers, that arrival matters a lot. A company that never looks at these prompts risks optimizing its site for a language the client almost never uses.
Note de quai. I keep three traces here: the phrase the model repeated, “clarify an industrial offer”; where it slipped, an editorial studio read as marketing consultancy; and the source that could help it, a proof page connecting the client prompt, the written deliverable and the boundary of the trade. The quay is no noisier for that. It simply carries better directions before the next passage.