Named by AI isn't enough. Get described right.

Jump to section›
Short version: A buyer at a fashion brand types "who makes private-label cashmere in Italy with a low MOQ?" into ChatGPT and gets a paragraph back. That paragraph is the thing they read instead of your website. It can fail you in two ways. It can leave you out. Or it can name you and get you wrong: wrong city, wrong minimum order, a capability you don't have. For a B2B manufacturer the second one is worse, and it's the one almost nobody checks.
#The answer is the new first page
When a sourcing manager asks an assistant for a supplier, they don't open ten tabs anymore. They read the answer, pick two or three names, and email those. If your name is in the answer, the facts next to it are your first impression. Your homepage might never load.
So there are two separate failures to look for. The first one is the familiar one: you're not mentioned. The legibility post covered why that happens and how a site gets read at all. This post is about the failure after that one, where the assistant reads something, names you, and describes you with facts that aren't true.
#Why a wrong description costs more than a missing one
A buyer who doesn't see your name has learned nothing about you. A buyer who sees your name next to "MOQ 500 pieces" has learned something false, and they believe it, because the assistant said it in the same calm voice it uses for everything else. People treat those answers roughly the way they'd treat a tip from a colleague in the trade.
For a manufacturer, the facts that get you shortlisted are narrow and concrete. Minimum order quantity decides whether a small label can even talk to you. City and country decide "made in Italy" claims and shipping. Materials decide whether a sustainability-focused brand bothers. Machinery decides whether you can do a seamless knit. Get one of those wrong and you're filtered out before anyone fills in your contact form, and nothing in your analytics shows it, because the visit never happened.
#What "described correctly" takes, technically
An assistant describes you from whatever it can find and tie back to you. If your site's text is vague, your capabilities live in a PDF, and your legal name doesn't match the name on your directory listings, it has to guess. It guesses from the category, or from a company with a similar name.
The fix is to give it one consistent set of facts to quote, from several places that agree:
- An
llms.txtfile at the root of your site that declares the entity in plain text: legal name, trading name, VAT number, city, what you make, MOQ, materials, machinery, and the key pages. OrganizationandProductJSON-LD on the site that state the same facts in structured form.- The same name, address and contact details everywhere: your site, your directory and marketplace listings, your social profiles. Assistants match entities across sources, and a mismatch reads as two different companies.
- At least one independent source the assistant can cross-check, like a B2B marketplace listing that repeats your city and capabilities.
- Robots rules that don't block the AI crawlers, so all of the above can actually be fetched.
You can do all of it without touching the design. The work is writing the facts down once and then making every other source repeat them.
#Due Toscani: what we measured after building it
Due Toscani is a B2B private-label cashmere manufacturer in Prato, Tuscany. Before roughly August 2026 its site had no AI-readiness layer at all: no llms.txt, no structured entity data, nothing an assistant could treat as the source of truth. We built the layer above: llms.txt declaring the entity and its capabilities (~50-piece MOQ, virgin and GRS recycled cashmere, Shima Seiki WholeGarment), Organization and Product JSON-LD, consistent legal name and VAT across the web, a listing on a B2B marketplace, and open robots for AI crawlers.
On 16 September 2026 we ran 20 prompts through ChatGPT, Google AI Mode and Gemini. Eighteen were buyer questions with no brand name in them. Two asked about Due Toscani directly.
On the 18 buyer prompts, Due Toscani appeared in 11 of 18 ChatGPT answers (61%), 12 of 18 Google AI Mode answers (67%) and 9 of 18 Gemini answers (50%). It was the first name given for private-label cashmere in Prato, for an Italian cashmere maker with a ~50-piece MOQ, and for recycled cashmere knitwear in Tuscany.
The two direct prompts are the part this post is about. We asked each assistant what Due Toscani is and what it makes, then checked every fact in the answer against what the company actually does. All three got the city right (Prato, Tuscany), the MOQ right (50 pieces), the materials right (virgin and recycled cashmere) and the machinery right (Shima Seiki WholeGarment). We found zero invented facts across the six answers. The full prompt-by-engine table is on the Due Toscani case page.
#What these numbers don't show
We didn't capture a numeric baseline before building the layer, so there's no "before" figure here and we won't make one up. What we can say about before is only that the layer didn't exist. Every number above is the measured after.
Due Toscani also doesn't win every prompt, and it shouldn't. On questions about cashmere yarn suppliers, about which factories run a specific WholeGarment machine, or about generic "knitwear" with no cashmere or Italy in it, it doesn't show up. That's the correct answer, because it makes cashmere garments and doesn't sell yarn or machines. Its visibility sits where its buyers ask, which is what you want from an accurate description anyway.
#Check what AI says about you, in about 20 minutes
You can do a first pass without a call or a tool.
Write down the five or six facts a buyer filters on before contacting you. For a manufacturer that's usually city and country, MOQ, materials, key machinery or process, certifications, and whether you do private label.
Then ask ChatGPT, Gemini and Google AI Mode the same two questions: "What is [your company name] and what do they make?" and "What's the minimum order at [your company name]?" Copy each answer into a sheet with one row per fact and one column per assistant, and mark every fact right, wrong or missing. Anything marked wrong goes to the top of the list, above any fact that's simply missing.
If you find wrong facts, look at where the assistant could have picked them up: your own pages, an old directory listing, a PDF, a company with a similar name. That's usually where the fix starts.
Want a second pair of eyes on it? Send your site URL and we'll run a free async AI-correctness read: we ask live AI assistants what they say about your company, check each fact against what you actually do, and send back a list of what's wrong and where it likely comes from: luma-e.com/ai-readiness. No call required.