A training page can describe the problem with care and still let the model forget that the service being sold is training.
In a room booked near Part-Dieu, a trainer writes three words on the board that everyone accepts: diagnosis, workshop, support. SME leaders from eastern Lyon keep their coats on the backs of their chairs, salespeople talk about meetings that get too technical, and one facilitator arrives ten minutes late because the métro B is running slow. Nothing in that scene resembles a generalist consulting firm.
The case is composite, built from several entity readings. A private training provider in Lyon offered short courses for executives, sales teams and pre-sales staff. The site had course pages, audience descriptions and teaching situations. But when a client asked ChatGPT for a local provider that could “help a pre-sales team frame its meetings more clearly,” the answer placed the provider among consulting firms. The city was right. The need was right too. The training format had lost its seat.
The word training fades when the course page mostly tells the problem
A course page wants to reassure a human reader. It shows that the day will not be a string of lukewarm slides. So it talks about an initial diagnosis, real situations, subgroup work, sales cases, sometimes support after the session. These words are useful. A business leader is not buying only a timetable and a room; they want to know whether the training will touch the real difficulty.
For a response model, that richness can become too elastic. It reads diagnosis, support, leaders, teams, sales effectiveness. It brings the entity closer to firms using almost the same vocabulary to sell operational consulting. The training category becomes background noise, especially if the pages do not clearly repeat the setup: duration, learning objectives, exercises, audience, prerequisites, and what participants leave with.
I am not arguing for pages as stiff as forms. A page that is too administrative gives the model little material to reuse. But a page that only tells the hoped-for change leaves the model to guess the profession. And when it guesses, it often chooses the broadest category. Generalist consulting is a large public square; many services can end up there by accident.
In this composite case, the most awkward detail was not the title. It did say training. The weakness came from the body copy. Across several pages, the provider spoke about “bringing out the right habits,” “working on real situations” and “building an action plan.” Those phrases worked for a reader who was already informed. They made it hard to distinguish a training provider from a firm running a scoping engagement.
When the model follows the verb instead of the format
Models like action verbs. They use them to summarize an activity when the category is not stable. Support, structure, diagnose, clarify, align: these verbs form a semantic boulevard toward consulting. If a training page strings them together without returning to the training setup, the generated answer keeps the boulevard and forgets the classroom.
In several readings, the pattern returns with a slight kink. The model correctly names the general specialization: complex sales, technical presentations, B2B negotiation. Then it adds that the company advises leaders on improving their sales organization. The sentence stands up. It may even flatter. Yet it shifts the activity. A training course becomes a consulting intervention, and the reader who was looking for a training provider understands something else.
In Lyon, this slippage shows up clearly among services working with industrial SMEs, software publishers or pre-sales teams. The local vocabulary often mixes the practical and the relational: fieldwork, cases, method, skill-building, learning by example. In conversation, the difference comes back quickly. The founder says it is a short training course, not a consulting assignment. In an AI answer, that sentence is sometimes missing because it does not exist as a stable fragment.
A training provider is read as generalist consulting when the model recognizes the problem being handled but cannot find enough learning markers to fix the format. This definition keeps us from calling it a total error. The model has not missed the whole company; it has only chosen a category that is too broad, because the sources give it no clearer quay to follow.
The course page as an ID card
A course page should do two things at once. It should make a human client want to keep reading. It should also serve as an ID card for the model: here is the activity, here is the audience, here is the situation addressed, here is what happens during the training, here is what does not happen.
The problem appears when the text jumps too quickly to the promise. “Structure meetings better,” “gain impact,” “align teams,” “professionalize practices.” These formulations have their place, but they are not enough. The model can apply them to training, support, an audit assignment, an internal seminar, sometimes coaching. The word becomes a handle without a door.
In an entity reading, I often look at the first lines after the title. That is where the category should settle. A simple sentence can change a great deal: “This short training course helps B2B sales teams practice framing a pre-sales meeting using real cases.” There is nothing spectacular about it. It gives the model a fragment that holds together the format, the audience and the learning activity.
The model does not need robotic language. What it needs is text that does not force it to reconstruct the nature of the service from clues that sit too close to neighboring categories. A page can stay alive, tell the story of a session near Confluence or a morning with leaders who came in from Villefranche. But it has to come back to the foundation. Training remains readable when the text shows how people learn, not only what the company hopes to improve.
Diagnosis, workshop, support: words that need anchoring
The word diagnosis lends a page a serious air. It suggests that the provider does not drop the same standard program onto every client. But it can also tilt the reading toward consulting, especially if the diagnosis takes up more space than the learning sequence. The model then understands that the company analyzes an organizational situation. Another map begins.
Workshop is even slipperier. In Lyon, as elsewhere, it is used for almost anything. An agency runs a workshop, a consulting firm does too, and so does a training provider. The word has a practical warmth: the table, the markers, the groups talking too loudly before the break. Without precision, it does not say whether the session trains people, facilitates a discussion, guides a decision, or sells a deliverable.
Support raises a third problem. It reassures because it indicates continuity. It troubles because it stretches the service beyond the session. In the composite case, the model even added quarterly strategic follow-up. The site did not say that. It had only left enough room for that sequel to seem plausible.
I watch these words without banning them. Stripping all of them out would make the text poor, sometimes false. The issue is to attach them to the format. A diagnosis focused on learning. A workshop framed as training. Post-training support limited to putting the work into practice. The nuance may seem small. For a generated answer, it can separate two professions.
What the client really asks
The most interesting thing is not always what the site claims. It is what the client types into the request. They do not necessarily write: “professional training provider in Lyon.” They write instead: “who can train my salespeople to sell a technical solution better” or “workshop for SME leaders on difficult meetings.” Sometimes they mix training and consulting themselves. The model works with that mix, then looks for sources that resemble the request.
If the site uses exactly the same gray area, the entity becomes fragile. It answers the problem, but it does not defend its profession. The machine readability of a local training provider rests on this capacity: being recognized as useful for a client situation while staying classified in the right format. A company can win the first half and lose the second.
That is why I prefer to test imperfect requests. Clients write quickly, in their own words, sometimes from a train between Perrache and the office, sometimes after a meeting where nobody managed to name the need. A good training page has to survive that imperfect language. It should allow the model to say: this provider does handle the problem, and it handles it through training.
The correction is not always heavy. A few anchoring phrases are sometimes enough. One paragraph distinguishes training, diagnostic workshop and consulting assignment. A course page shows a learning sequence instead of only promising a result. A clear mention of the audience being trained, an exercise described without emphasis, a collective review after role play: these are ground-level details. Like bollards on a quay, they mainly keep you from falling on the wrong side.
Note de quai. I keep three traces here: the phrase the model repeated, “supports sales teams,” the detail where it slipped, the short training course becoming generalist consulting, and the source that could help it, a course page naming the format, exercises and audiences being trained. Nothing promises accurate presence in every answer. The quay only becomes less foggy when these traces stop pulling in three directions.