Raw data from an original study measuring which luxury jewelry brands ten large language models recommend when asked the same question.
Canonical version of the analysis: https://websem.ro/resurse/aeo/studiu-bijuterii-lux-romania English: https://websem.ro/en/resources/aeo/study-luxury-jewelry-romania
This repository/record holds the data only. The full write-up, charts and interpretation live at the canonical URL above. Please cite that page.
One prompt, word for word, put to ten public LLMs:
Which luxury jewelry brands from Romania do you recommend for wedding bands and engagement rings? Give me a top 5, with a short argument for each and the sources you rely on.
It contains no brand names. Every brand in this dataset was chosen by the model itself.
| Parameter | Value |
|---|---|
| Query date | 13 July 2026 |
| Models queried | 10 — ChatGPT, Claude, Gemini, Copilot, Grok, Perplexity, DeepSeek, Kimi, GLM, Qwen |
| Runs per model | 1 (no regeneration) |
| Session | no history, no memory, no custom instructions |
| Brands named in the prompt | 0 |
| Available positions | 50 (10 models × top 5) |
| Distinct brands returned | 29 |
| Localisation | Romania, Romanian language |
- 29 distinct brands across 50 slots. Ten models, asked identically, did not converge on a shared answer.
- No brand appears in all ten lists. The widest coverage is TEILOR, named by 6 models out of 10.
- 19 of the 29 brands (66%) were named by exactly one model. Two thirds of the category is visible to a single assistant and invisible to the other nine.
- Average pairwise agreement between models: 18%. Pick any two models and they share, on average, 0.91 of their five slots.
- Only five brands ever reached first place in any model's list.
- One model is an island. Copilot's top 5 overlaps almost not at all with the other nine — it returns a separate set of brands entirely.
| File | Rows | Contents |
|---|---|---|
llm_rankings_raw.csv |
50 | The raw result set. One row per model-position pair: every brand every model returned, in order. Start here. |
brand_visibility.csv |
10 | Aggregate leaderboard, derived from the raw file. Contains only brands named by 2+ models. |
bijuterii-clasament.png |
— | The leaderboard chart |
CHECKSUMS.txt |
— | SHA-256 of each CSV |
Important: brand_visibility.csv is a summary and is deliberately partial —
it excludes the 19 single-model brands. Since "most brands are seen by exactly
one model" is the study's central finding, any analysis should start from
llm_rankings_raw.csv, not the aggregate.
Column-by-column descriptions: see DATA-DICTIONARY.md.
brand_visibility.csv derives entirely from the raw file:
import csv
from collections import defaultdict
rows = list(csv.DictReader(open("llm_rankings_raw.csv", encoding="utf-8")))
by_brand = defaultdict(list)
for r in rows:
by_brand[r["brand"]].append(int(r["position"]))
print(len(rows)) # 50
print(len(by_brand)) # 29
print(sum(1 for v in by_brand.values() if len(v) == 1)) # 19
for brand, positions in sorted(by_brand.items(), key=lambda kv: -len(kv[1]))[:3]:
print(brand, len(positions), round(sum(positions) / len(positions), 1))
# TEILOR 6 2.0 · Malvensky 5 1.2 · ...Declared up front:
- One prompt, one run per model. This is a snapshot, not a time series. Episode 2 of the series switched to a repeated panel for exactly this reason.
- 10 models, not every assistant on the market.
- A single user profile. No history, no personalisation, queried from Romania in Romanian. A logged-in user may get different answers.
- Models are non-deterministic and update continuously. The figures describe model behaviour on 13 July 2026, not a permanent truth.
- Mentions are not sales. This measures presence in answers, nothing more.
This study presents results obtained by querying 10 public artificial-intelligence models on 13 July 2026, using a single question, identical for all, which contained no brand name. All brands mentioned in this material were named by the AI models, not by the study's authors.
The material is market research and information. It does not constitute an evaluation of the quality of products, services or commercial performance of any brand, does not express a Websem opinion on these aspects, and has no comparative-advertising purpose.
The names and trademarks mentioned belong to their rightful holders and are used exclusively for identification, within a factual analysis, in accordance with fair industrial and commercial practice.
The results reflect the state of the models on the query date and may change. Any trademark holder may request a factual correction or right of reply, which will be published.
Contact for corrections: office@websem.ro
Creative Commons Attribution 4.0 International (CC BY 4.0).
Alexandrescu, D. C. / Websem (2026). AI visibility of Romania's luxury jewelry, July 2026 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21724399
DOI (all versions, resolves to the latest): 10.5281/zenodo.21724399 · Canonical analysis: https://websem.ro/resurse/aeo/studiu-bijuterii-lux-romania
Machine-readable citation metadata: CITATION.cff.
Author: Dan Cristian Alexandrescu — ORCID 0009-0007-1994-6700 · Websem (SUPREMIUM GENESIS S.R.L.)
This is episode 1 of an ongoing series measuring how Romanian markets appear in AI answers. Episode 2 covered the book market across 5 AI engines, with a repeated weekly panel: https://websem.ro/resurse/aeo/studiu-piata-carte-romania (DOI 10.5281/zenodo.21736211)
Websem — AEO/GEO agency, Bucharest. https://websem.ro