The same question — “Que tipo de empresa é esta e para quem serve?” — produced three answers for Object A that looked close at first glance: a customer messaging platform, a CRM system, a marketing service.
Object A is a composite scenario: a small Portuguese B2B platform for customer communications that sells its product to service companies outside the local market. In all three answers, the model kept the B2B context and the customers. In one run, it stayed almost entirely inside the frame of service communication. In another, the first major neighbor was a CRM system. In the third, campaigns and potential customers appeared nearby. The prompt did not change. The answer’s path did.
For Object B — a composite scenario about an operational B2B platform for small tour operators and local excursion services — the same fixed question moved differently. In one run, the model described a tool for local tours, bookings, and incoming requests. In another, it drifted toward hotel software. In a third, it left a cautious formula about tourism operations, with almost no role for the small team. The words were tidy. The map quietly shifted each time.
One Prompt Is One Run
A user has a reasonable expectation: the same question should return the same picture. In conversation with a model, that expectation becomes even stronger because the answer looks whole. It does not show the list of sources and forks behind it. It speaks as if the reader is seeing an already assembled note.
The laboratory works with a more cautious unit of observation. One prompt is a concrete run, not the model’s settled position on a brand. The answer is shaped by the model, mode, dialogue context, possible search layer, language, order of follow-up questions, and the ambiguity of the question itself. Even when the prompt string matches, the assembly can move by different routes.
In the series for Object A, the beginning often looked safe: B2B, customers, communications. The difference appeared at the next turn. If the model chose a CRM neighbor, the answer moved toward contact management and a sales logic. If a marketing neighbor entered the answer, the product began to sound like a tool for campaigns. If the answer held onto the service team, the function stayed closer to the original scene.
For Object B, the repeated question was not only testing industry knowledge. The model almost always saw tourism. The question was where it took tourism next: toward local excursions, toward accommodation, or toward an empty formula about operations. So the same prompt does not show a stable brand profile. It shows a set of possible roads around it.
The Semantic Path Matters More Than Matching Words
A semantic path is the direction in which an answer leads a company across repeated runs. The words may change, but a recurring route toward one category, function, or error becomes meaningful. This definition matters for Atelier das Entidades because literal repeatability in AI answers is often deceptive.
Two answers may share very few phrases and still make the same substitution. For Object A, “customer relationship tool,” “communications platform,” and “customer engagement solution” sound like different formulas. If all of them move the product away from service operations and toward sales or marketing, the path is one. The wording changed. The shift remained.
For Object B, a similar picture appears through bookings. One answer speaks about reservations, another about customer requests, a third about managing a tourism business. The laboratory does not look only at vocabulary. It looks at the scene built after those words. Is there a small tour operator, a guide, tour time, multilingual description? Or has the model moved to a hotel system, where booking and guest mean a different kind of work?
This is why one successful answer does not settle the question of AI visibility. It may be successful because that particular run chose the right route. In the next run, the same prompt may open the neighboring door. That is not a verdict on the brand, but it is a signal: the trace is not stable enough if the model keeps choosing again between several nearby shelves.
Where the Road Splits
The split often begins in the first neighborhood after a neutral description. An answer about Object A may begin the same way: Portuguese B2B platform, customers, communications. Then one word lays down the rail. CRM leads to contacts, sales, and pipeline. Marketing service leads to campaigns and acquisition. Operational communication leads to messages, context, and handoff between employees.
For Object B, the rail is laid by what the booking is for. If the model holds the excursion, the booking connects to route, time, and the team’s task. If it chooses the hotel shelf, booking turns into accommodation. If it drifts into general tourism operations, the product loses the working scene and remains almost without a function. All of this can happen in answers to the same question.
These divergences should not be read as simple “model randomness.” It is more useful to ask which parts of the AI trace behave like forks. For Object A, the fork appears after words about customers and communications. For Object B, after bookings and operations. If several runs keep choosing different roads in the same place, that already gives material for editorial and methodological analysis.
The laboratory therefore records more than the final category. It records the order of the answer. What appears first after the name? Which task is named as the main one? Which audience is implied? Where does the model begin to speak in generalities? Order matters because the early neighbor often sets the remaining comparisons. The answer seems to build a bridge: the first plank is still correct, then the bridge heads toward another shore.
A Map of Entity-Loss Routes
Atelier das Entidades classifies four ways a model loses a brand’s entity: it shifts the category, substitutes the function, attracts a neighbor, or leaves a blank space. In the material about a repeated prompt, this typology works as a map of routes between answers.
A category shift is visible when Object A turns from a customer communications platform into a CRM system or marketing service. Neighbor attraction appears slightly earlier: the model places a close category nearby because the language of customers, messages, and automation is familiar to it. Function substitution happens when communication with customers begins to be described as acquisition, segmentation, or a campaign. A blank space appears in answers where the model writes “B2B solution for customer operations” and does not show the process.
For Object B, the same four routes look different. The category shifts toward hotel software. The function changes from coordinating local tours to managing accommodation or general support. The neighbor becomes a hotel system, travel agency, or customer request service. The blank space appears in the formula “solution for tourism businesses” when there is no excursion booking, customer question, team task, or multilingual description behind it.
Such a map helps avoid collapsing different errors into one complaint about “instability.” One answer may be inaccurate because of category. Another because of function. A third because of emptiness where the audience should be. For working with an AI trace, these are different tasks. Category is held by one set of links, function by another, and blank spaces are filled with scenes of use.
How to Read a Safe Run
A safe run usually tells part of the truth. It names the country, language, B2B context, broad industry, and sometimes repeats a strong formula from the original description. The reader sees familiar words and relaxes. But a repeated series shows that a correct beginning does not yet guarantee the right road.
The laboratory therefore reads an answer by direction. Where does it lead the brand after the first sentence? Which neighbors appear unasked? What became the product’s main action? Which audience surfaced on its own? For Object A, a safe answer should keep the service team and operational customer messages. For Object B, it should keep the small tour operator, local tour, booking, incoming question, and internal task.
One successful run is useful as an observation, but weak as evidence of stability. If the next answer sends Object A toward marketing and Object B toward a hotel system, the model has not anchored the entity firmly enough. It is not necessarily making a crude mistake. It is choosing one of the available routes, and some of those routes lead into a neighboring market.
For a brand, this changes the practice of checking. A single question should be replaced by a short series of repeated runs with the same wording. The important thing is not literal sameness across answers, but the recurring path. If the divergences are different every time, the conclusion should stay cautious. If they keep returning the brand to the same neighbor or blank space, the observation becomes denser.
Limits of the Observation
This material does not claim that the same prompt will always produce different semantic paths. Under some conditions, answers may be stable: if the prompt is narrow, the context is rich, and the model is running in a mode with tighter parameters. The laboratory looks at practical user scenarios, where the question is often broad and asks the model to explain the company type, audience, and purpose of the product.
A fixed-prompt series does not replace a full audit of a brand’s open texts. It shows model behavior under specified conditions, but it does not reveal the whole history of the error. It is not possible to say with confidence which exact document, translation, catalogue, or site phrase led Object A toward a CRM system and Object B toward hotel software. What can be recorded is the answer, its distance from the source description, and whether the route repeats.
Other prompt formulations matter too, but they belong to a separate layer of checking. A question about alternatives, a supplier-choice question, and a question about product function may produce other forks. The subject here is narrower: the same prompt across several runs. Under these conditions, Object A and Object B did not always preserve the connection between category, audience, and practical function. That becomes an observation, not a final diagnosis of the brand.