An English assistant answer does not become accurate by translating goodwill. It becomes accurate when the same business facts are visible enough to survive the trip from Italian page to English summary.
The first English answer looked polite, which made it worse. The assistant described a small workshop in Emilia-Romagna as “a boutique furniture label known for refined interiors.” The real business, in the composite scenario I am using here, was a fourteen-person family workshop making custom wood fittings for boutique hotels, restaurants and private renovations. It did not sell a furniture collection. It did not design full interiors. One sentence had dressed a practical supplier in a nicer, wronger coat.
The Italian site was not empty. It had photographs, a warm Chi siamo page, some product descriptions, and a contact page with the province visible near the footer. The English page, though, had been cut down to a few graceful lines: “Italian craftsmanship for refined spaces.” That was the sentence the assistant loved too much. It had rhythm. It had no working evidence.
The English answer is often built from the thinnest plank
When owners ask me why an English-language assistant answer is poor, they often begin with the Italian site. “But everything is there,” they say, and sometimes they are half right. The Italian pages may carry useful facts: the workshop’s region, the type of customers, the made-to-measure nature of the work, the materials, the service limits, the difference between supply and design. Then the English page takes only the mood and leaves the evidence behind.
An assistant can translate, but translation is not the same as entity recovery. If the English prompt asks, “Who makes custom wood fittings for hotels in Emilia-Romagna?” the model may search, retrieve snippets, compare pages, and assemble a short answer. If the business has an English page with only atmospheric wording, the assistant has to decide what the entity is from fragments. A thin English page behaves like a plank laid across a canal: it may hold a human reader for a moment, but it is not enough for a machine carrying a heavy basket of assumptions.
The recurrent pattern is simple. Italian pages describe the business with local familiarity, because the owner assumes the reader already understands the category. English pages describe the business with aspiration, because the owner wants to sound credible to foreign buyers. Between those two habits, the assistant loses the plain facts. It may keep the country and the beautiful adjectives. It may drop the region, the buyer type, the difference between a workshop and a retailer, and the service boundary that prevents a wrong category.
A thin Italian site can also create weak English answers even when there is no English page at all. In those cases, the assistant may translate snippets from Italian, rely on platform descriptions, borrow language from image alt text, or use directory categories. The result is often acceptable in grammar and unstable in substance. It sounds like somebody describing a street after seeing only a shop sign and two reflections in a window.
What “thin” means when a machine reads the page
Thin does not always mean short. I have seen long pages that are thin because they repeat the same soft phrase ten times. I have also seen compact pages that are strong because every sentence does a job. Length is secondary. The real question is whether the page leaves enough visible facts for an assistant to repeat without inventing connective tissue.
For my work, a thin bilingual evidence page is a page where the assistant can see the business mood but cannot reliably answer what the business is, where it works, who it serves, and what it does not provide. That is the working definition I use, because the missing facts show up directly in assistant errors. A page may feel elegant to a human and still be thin for retrieval if its category, location, customer and service boundaries are left to implication.
This is where Italian-to-English pages become fragile. Italian business writing often tolerates a certain closeness of context. A phrase like lavorazioni su misura may be clear to a local buyer who knows the workshop, the region and the material from conversation. In English, “custom work” may float anywhere: furniture, fittings, restoration, design, installation, even luxury décor. The assistant has no cousin in the village to ask. It reads the words it gets.
In the composite workshop case, the Italian page used several phrases that made sense in conversation: “projects for spaces with character,” “solutions for hospitality,” “wood details for interiors.” The English version compressed these into “bespoke interiors.” That phrase is dangerous. It can point to a design studio, a furniture brand, a contractor, an architect’s collaborator, or a maker of custom fittings. A human buyer may click through and understand. An assistant may answer before that understanding arrives.
I call this problem the translated fog layer. It appears when the Italian page contains some entity facts, but the English page keeps only atmosphere, prestige and broad category hints. The fog is not created by bad English. Often the English is smooth. It is created by the removal of the dull words that machines need: workshop, custom wood fittings, Emilia-Romagna, hotels, restaurants, private renovations, made to order, not a retail furniture catalogue, not a full interior-design studio.
The minimum English evidence is smaller than a full translation
Owners sometimes think the repair must be expensive and heavy: translate the whole site, rebuild the menu, write a second brand voice. Sometimes that is useful. More often, the first fix is more modest. The English answer needs a minimum evidence layer, not a literary twin of the Italian site.
A minimum evidence layer is a set of visible English sentences that carry the same entity facts as the Italian pages, because assistants need stable facts more than fluent decoration. That definition sounds dry, but it saves money. The owner does not need to translate every warm paragraph about family history before the assistant can describe the business correctly. The owner needs a few durable blocks that say the business name, category, region, customer type, main work, and limits in plain English.
For the workshop, the repair began with one paragraph placed on the About page and repeated in shorter form on product and contact pages. It named the entity as a family workshop in Emilia-Romagna. It said the work was custom wood fittings and built elements for boutique hotels, restaurants and private renovations. It separated making and fitting from full interior design. It mentioned that projects are made to order, not sold as a standard furniture collection. Nothing poetic. Almost boring.
Boring can be a kindness.
The English paragraph did not erase the Italian tone. The Italian page could still carry tradition, material knowledge and family continuity. The English page simply stopped asking an assistant to infer the skeleton from the perfume. After the repair, I would expect repeated assistant runs to have less freedom to call the business a furniture brand. I would still test this rather than declare victory from one answer. One clean answer is pleasant. Several similar answers begin to look like evidence.
The same principle applies to small hotels, repair trades, food producers, studios, and service companies. If the Italian page says the full truth and the English page says only “authentic experience,” the English answer will be poor. If the English page says the business plainly, even in five or six sentences, the assistant has something it can carry.
Translation can move facts out of view
The most damaging translation choices are often made with good taste. The owner or marketer removes repetition, softens blunt category words, and replaces practical detail with a phrase that sounds more international. This can improve the page for a certain human reader and damage it for an assistant at the same time.
In page work I look for facts that disappear during translation. Region becomes “Italy.” A workshop becomes “brand.” Made-to-measure work becomes “bespoke lifestyle.” Repair becomes “care.” A service area becomes a mood. The Italian page says the company works with hotels and restaurants in a defined region; the English page says it creates spaces for discerning clients. Now the assistant has a polished coin with the face rubbed off.
There is also the opposite problem: the English page adds facts the Italian site does not support. I see this when a small business wants to sound broader for foreign customers. It may add “international projects,” “design consultancy,” “luxury hospitality,” or “sustainable interiors” where the Italian page gives no clear evidence for those claims. The assistant may then repeat the inflated version. Later, when an owner complains that the answer is exaggerated, the cause is sitting quietly on the English page.
A careful bilingual page does not need identical music in both languages. Italian and English can sound different. They should not carry different entity facts. If the Italian page says the company serves restaurant renovation projects and the English page says it serves luxury residences, the assistant is not being mysterious when it blends them. It is reading a split personality.
For a practical check, I place the Italian and English pages side by side and mark the facts that must survive. I do not start with style. I start with the dull columns in my head: name, legal or trading identity when useful, category, region, address cue, service area, customer type, product or service limits, booking or buying route, and what the company is often mistaken for. When one of those facts exists in Italian and vanishes in English, the answer risk increases.
Platform English can overpower the business page
When the business page is thin, assistants often turn to platforms. For a guesthouse, that may be booking profiles and review sites. For a workshop, it may be marketplace listings, directories, supplier pages, image captions, or a distributor’s description. These sources are not enemies. They become too powerful when the owner’s own site fails to state the same facts more clearly.
The composite workshop had a directory profile that called it a furniture maker. That was not completely false. It worked with wood and produced built elements. But “furniture maker” was too broad and pulled the assistant toward product-brand language. Another page from a regional tourism feature used “artisan interiors,” probably because it sounded attractive in a short caption. The English site, being thin, did not correct either source. It nodded along.
In most cases, assistants prefer a path of least resistance. A visible platform sentence with a neat category may beat a vague business page with better truth. This is uncomfortable for owners because the platform was never meant to carry the full identity. It was meant to get clicks, bookings, inquiries, or a little visibility. Yet in an assistant answer, that small platform sentence can become the spine.
The repair is not to remove every outside profile. It is to make the official pages strong enough that outside profiles become supporting evidence. The business site should hold the cleanest statement of category and service boundary. Platforms can echo it, but they should not be the only place where the machine finds a usable noun.
For English answers, this means checking the platform language in English too. An Italian profile may be tolerable while the English version chooses a wider category. An automatic platform translation may turn laboratorio into studio, falegnameria into carpentry service, or fittings into furniture. Each small change adds weight to the wrong bucket.
Test the English answer as a pattern, not a mood
A single English prompt tells me little. It may catch the assistant on a good run or a lazy one. I test several prompts around the same entity: direct name prompts, category prompts, local supplier prompts, comparison prompts, and buyer-intent prompts. The point is not to celebrate a mention. The point is to find the repeated weakness.
In the workshop case, I would ask in English about custom wood fittings in Emilia-Romagna, suppliers for boutique hotel renovation, the business by name, and alternatives to larger Italian furniture brands for made-to-order hospitality work. I would then mark the exact sentence where the assistant drifts. Does it call the workshop a design studio? Does it omit the region? Does it imply a retail catalogue? Does it borrow scale from a louder competitor? The wrong sentence is more useful than the whole answer.
The repair should then be tested again, after the pages are live and crawlable where relevant. I do not promise instant correction. Generative systems vary, retrieval changes, and some assistants carry older traces longer than owners expect. Still, a clearer page gives future runs a better handle. It is like putting a metal nameplate on a drawer that everyone has been opening by guesswork.
The work is small, but not cosmetic. A thin Italian page or a thin English page changes what the assistant can safely say. The owner may think the page is only for human persuasion. In generative search, the page also becomes a witness. A weak witness gives charming answers and poor testimony.
The Vellumari Margin — Name on the page: an Italian SME with English buyers must not let its English page keep only atmosphere. Wrong shadow: the assistant may turn a workshop, guesthouse or service firm into a broader lifestyle category. Clean line: carry the same entity facts in English, even if the tone changes. Trace to leave: repeat category, region, customer type and service limits on the About, product or service, and contact pages.