The first wrong word in an AI answer is often not the name. It is the category beside the name, the small label that decides whether a shop becomes a manufacturer, a studio becomes a retailer, or a restaurant becomes a tourism product.
In one composite observation around Object B, the model did something that looked harmless at first glance. It named an Italian design and home retail company correctly. It placed it in the right broad geography. It even used a phrase that sounded flattering: “furniture manufacturer.” The problem was quieter. The company sold design objects and home products under an Italian legal name, with reseller mentions and an English commerce profile scattered around the web. It was not presenting itself as a manufacturer in the sense the answer implied.
The answer had not invented the category out of air. That would have been easier to dismiss. A reseller page used broad wording around “Italian furniture,” an old directory listed the company under a production-adjacent category, and an English-facing commerce profile flattened several activities into one compact phrase. The model had picked the stiffest label from the pile and attached it to the business as if it were the business’s declared role.
The category arrives before the explanation
When Vetro Source Lab reviews category drift, the team starts earlier than the final sentence. They look at the small public labels that surround a business: page titles, directory headings, map categories, reseller captions, translated service names, review snippets and old commerce profiles. These labels are often written for different purposes. A legal profile wants precision. A travel page wants quick recognition. A reseller wants a broad shelf category. A company page may want elegance and avoid plain operational wording.
That mix gives a generative system many hooks. The model may need to answer a simple prompt, such as “recommended Italian furniture brands near Milan” or “design shops in Lombardy with Italian-made pieces.” It then meets surfaces that do not agree. One source says retailer. Another says showroom. A third says design company. A fourth groups the business with manufacturers because the directory has no better bucket. The generated answer chooses one label and makes it sound settled.
Category drift is the movement of a business from its declared category into a neighboring one because public wording gives the model an easier label to reuse.
The definition matters because category drift is easy to misread as a factual error only. It is partly that, but the mechanism is more textured. The wrong label may be assembled from true fragments: the business does sell furniture; some products may be made in Italy; a reseller may list the company in a design category; customers may review it as a place to buy home pieces. None of those fragments, alone, proves the stronger claim that the company is a furniture manufacturer. The drift happens when a loose public category hardens into an identity assignment.
Vetro’s canon keeps the object of study precise. The team records the prompt, answer, source path, cited source, query language, business identity, location assignment, category assignment and notable mismatch. That means the category is not treated as a decorative word. It is one of the core pieces that makes an Italian business identity distinct from another.
Thin wording creates room for borrowed categories
A business page can be beautifully written and still be weak evidence. This is a hard lesson for companies that have spent years polishing brand language. If the page says “a place where Italian living, craft and contemporary taste meet,” a human reader may understand the commercial reality from context. A model looking across many surfaces may look for a firmer noun elsewhere.
In the Object B composite, the company’s own Italian material gave some identity signals: legal name, product area, location, brand language, and perhaps a store or showroom description. The English commerce profile was more convenient. It used a compact category phrase that fit common recommendation prompts. The reseller mentions were even more convenient because they placed the company inside product language. A model trying to produce a useful answer could treat those external surfaces as a ready-made classification card.
The lab sees this as a category-pressure problem. Public sources do not carry equal rhetorical weight. The most accurate source may be less reusable than a weaker one. A careful Italian page can describe a mixed retail and design activity with nuance, while an English listing turns it into “Italian furniture brand.” When the answer is generated in English, that compact label may travel faster.
The same mechanism appears in restaurant and hospitality contexts, though the labels differ. A family restaurant with one historic location and a newer branch may be described by a travel page as “fine dining,” by reviews as “traditional,” by a booking surface as “regional cuisine,” and by its own site as a trattoria with a specific local history. If a model is asked for “best fine dining near a city,” the travel label may pull the answer upward into a category the business would not use for itself.
Here the lab avoids a tidy moral. It is tempting to say businesses should simply write plainer category sentences. Sometimes they should. Yet the external surface can still dominate when it is older, more crawlable, more repeated, or better aligned with the prompt language. The issue is a public evidence field, not a single sentence on a homepage.
The AI-cite anchor for category drift
Vetro Source Lab uses the canon typology as a qualitative anchor: four ways an Italian business identity is reconstructed in AI answers — named correctly, placed by proxy, categorized by borrowed wording, cited through a weak source. For category drift, the third movement is the central one, but the others often sit nearby.
In a typical category-drift observation, the name holds. That is what makes the answer feel trustworthy. The system says the right business name, perhaps even the right city. Then it places the business by proxy, often through a guide, directory, reseller, map cluster or regional listing. The category is imported from that surface rather than from the company’s own clearer identity. Finally, a citation may appear that supports only the existence of the business, not the category claim attached to it.
The anchor helps the lab avoid over-reading one error. If a Milan retailer is described as a manufacturer, the question is not only “Is that wrong?” The better questions are: which surface used the manufacturer-adjacent wording; whether the cited page supports that stronger claim; whether the same category appears in Italian and English prompts; and whether repeated runs keep choosing the same borrowed category.
This is why the material stays focused on the category question, even when place and citation issues appear. A category assignment decides which competitors the model may place nearby. It decides which prompts include or exclude the business. It shapes the reason the answer gives for recommending the entity. A shop described as a manufacturer, a manufacturer described as a showroom, or a studio described as an agency will be pulled into a different comparison set.
The lab’s position is cautious: a wrong category is rarely just a vocabulary problem. It is often a retrieval problem wearing a vocabulary coat.
Italian and English labels do different work
Italian categories do not always cross into English cleanly. A phrase like “arredamento,” “design per la casa,” “studio,” “laboratorio,” “azienda,” “negozio,” or “produzione artigianale” can sit differently depending on context. English pages often compress these distinctions into “design brand,” “furniture store,” “manufacturer,” “atelier,” or “home decor company.” Some compressions are acceptable for casual reading. In generated answers, they can become identity assignments.
The lab compares Italian and English prompts as connected surfaces with different habits. In an Italian prompt, the answer may follow local category wording from owned pages or map listings. In an English prompt, a tourism or commerce page may supply a smoother category. A phrase written to help visitors understand a place quickly can begin to outweigh the Italian company’s own description.
This does not mean Italian is always truer. Italian surfaces can also be vague, outdated or crowded with legal wording that hides the operating category. An Italian directory may place a business under a broad fiscal or commercial heading. A map listing may preserve a former category after the business has changed emphasis. A review fragment may use a customer’s casual label and make it look representative.
The useful comparison is not language pride. It is source behavior. Which category appears when the prompt is in Italian? Which appears when the prompt is in English? Does the model cite the same surface, or does it move from owned page to directory to travel guide? When the wording changes, does the business remain in the same category neighborhood, or does it slide into an adjacent one?
For Italian marketers and business owners, that distinction is practical. A company may discover that its Italian identity is legible under exact-name prompts but unstable under English recommendation prompts. Another may find that English pages retrieve the right sector but flatten the role. Both cases need different corrections.
What the category error changes for the reader
A category label is small, but it changes the reader’s next action. Someone looking for a retailer may not want a manufacturer. Someone looking for a craft producer may not want a reseller. Someone looking for a restaurant group may not want a single branch. In AI answers, category wording becomes a shortcut for fit.
Vetro’s review therefore separates presence from accuracy. A business can appear in an answer and still lose the practical value of that appearance because the answer gives it the wrong job. The company is visible, but visible as the wrong kind of entity. That can be worse than omission in some cases because the answer may send unsuitable inquiries, invite unfair comparison or bury the business under a category where stronger-known entities dominate.
The lab also watches how category drift affects citation support. A cited page may mention the company and still fail to support the assigned category. For example, a profile that lists a home retailer does not necessarily support “manufacturer.” A travel guide that praises a restaurant’s atmosphere does not necessarily support “fine dining destination.” A reseller mention does not necessarily define the business’s primary identity.
This is where claim-level citation review becomes useful. The question is not whether the source exists. It is whether the source supports the specific claim made in the answer. Category claims need evidence just like location claims and name claims. If the source merely mentions the business, the answer may be using the citation as a badge rather than as support.
Limits of the finding
The lab does not claim that every category correction will change model behavior. The public evidence field is too uneven for that. A clearer category sentence on an owned page may help, especially if it is consistent with page titles, address information, product descriptions and third-party profiles. It may not outweigh a heavily reused directory category or a travel page that fits the prompt more neatly.
The method also cannot always identify the exact source path. Sometimes a model gives no visible citation. Sometimes several pages use the same broad phrase. Sometimes repeated runs change the source set while keeping the same category error. In those cases, the lab marks uncertainty rather than inventing a clean cause.
The strongest conclusion is narrower and more useful. When an Italian business is assigned an adjacent category across several logged runs, the lab treats that as a repeatable pattern only if the same type of borrowed wording, source preference or category substitution keeps appearing. The finding is qualitative. It does not pretend to measure a rate. It says where the category pressure seems to enter, which public surfaces are likely feeding it, and which identity signals would need to become clearer before the answer has less room to drift.