The AI visibility of Romania's luxury jewelry

Microsoft Copilot recommended five brands — Rosental, Adorri, Jasmin, Perle și Pietre, Diamondstone — none of which appear in any of the other nine lists. Not TEILOR, not Malvensky, not Sabion. Zero overlap.

The reason shows in its arguments: Copilot cited a rating (“Jasmin — excellent reputation, 4.93 out of 906 reviews”), which gives away the source. Copilot answers from the live Bing index, anchored in SEO-optimized pages and review aggregators. It doesn't answer from “what it knows about Romanian luxury” but from what ranks well in search right now.

Consequence for brands: Copilot is the only one of the ten you win with classic SEO and review volume. For the other nine, SEO isn't enough.

TEILOR has the widest coverage (6 of 10 models), but an average position of 2.0 and a 5th place in Qwen's list. Malvensky appears in only 5 models, but ranks first in four of them, with an average position of 1.2.

TEILOR has category authority — it's the name that surfaces when the model enumerates the market. It comes from scale, a store network and, above all, financial press (Ziarul Financiar, Forbes wrote about revenue and a potential IPO). Source-anchored models cite figures about TEILOR because the figures exist in print.

Malvensky has entity salience — it's the name that surfaces when the model has to say what Romanian luxury means. It comes from a single, verifiable, repeated fact: “Official Supplier to the Royal House of Romania.” ChatGPT and Qwen cite it almost identically.

The lesson: one distinctive, verifiable fact is worth more than ten pages of adjectives.

Claude places Sabion's workshops in Târgu Mureș (citing Profit.ro). Gemini places them in Bacău. One model cites the source, the other doesn't. This isn't a difference of opinion but a factual error in one of the models — and it perfectly illustrates why brand data must live in citable sources: where the model can't find, it invents plausibly.

Across the 10 lists appear, separately: Stil Diamonds (Grok, Kimi), StilUnic (DeepSeek) and, in Claude's answer, a mention of a “Stil Unic” workshop on Calea Victoriei — flagged explicitly as an article marked “paid content” and therefore excluded from its top.

Three entities, nearly identical names, one category. For a model, disambiguation is hard and the result is fragmentation: no variant gathers enough critical mass to reach the top across several models.

Claude was the only model to explicitly flag that part of the public information about Romanian jewelry is brand communication and advertorial, not independent review, and excluded a source because it was marked “paid content.”

The other models didn't make this distinction. The commercial takeaway is uncomfortable but real: paid PR becomes, over time, an “AI recommendation” — for most models. The difference between an advertorial and a news story dissolves in the training corpus.

Sabrini appears in 3 of 10 models — Grok, Kimi and Qwen. It appears in none of the models that anchor their answer in web search with source citation (Claude, Perplexity, Copilot). Qwen, the only one to rank it high (2nd), cited its sources: facebook.com and instagram.com.

This is the study's most instructive correlation. A brand whose public presence is predominantly social will be recommended by models that feed on a broad corpus — and ignored by engines that demand confirmation from independent editorial sources. Social media gets you into AI. The press keeps you there.

The most elaborate arguments in the whole study belong to GLM — describing techniques (“invisible setting”), histories (“founded in 1875”) and sources (“Harper's Bazaar Romania,” “Forbes Romania,” “Ziarul Financiar”), with no links. Four of GLM's five brands appear in no other list. DeepSeek lists, under “Source,” the brand's own name — i.e. not a source.

Rule of thumb: the less verifiably a model cites, the more likely the recommendation comes from generalization, not data. The beauty of the argument is inversely proportional to its grounding.

19 of the 29 brands appear exactly once. For each of them, AI visibility is a lottery: if your future customer asks a different model than the one that knows you, you don't exist.