Luxury jewelry — visibility in AI answers · edition 2026-07

Each row names the field of the edition data that will not support a comparison, and the exact comparison it will not support. The list sits here, in plain view, because this is the only place where it can still change anything.

• market.tier

  Everything below is tier “pilot” and enters no aggregation in the index: not the index totals, not a brand ranking, not a comparison between markets. The edition is published so it can be read and checked, not so it can be compared.

• method.design

  A different instrument. Here: one question, asked once, of ten assistants. In tier A editions: a panel of purchase questions, run across five surfaces from three providers. The two do not measure the same thing and do not belong in the same table.

• brands[].mentions

  The figures in the mentions column are slots, not free mentions — `mentionsAreSlots` is true. Every model had to name exactly five brands, so a brand can appear at most once per answer and the total is fixed at 50 (10 answers × 5 positions). In a tier A edition a mention is an event; here it is a position in a list that was demanded.

• brands[].weighted

  The position-weighted sum is not stored — `weighted` is null on every row — and no position score can be computed for this edition. The reason is arithmetic, not editorial: with a forced top 5, 100 + 50 + 33.33 + 25 + 20 = 228.33 comes out identical for each of the ten answers (verified: all ten give 228.3333, total 2,283.33). It would be a property of the prompt presented as a result. The distribution stays in `positions`, for anyone who wants to redo the arithmetic.

• surfaces

  Ten assistants fit into seven surface rows, because three of them — Kimi, GLM and Qwen — have no code of their own in the schema and were pooled with DeepSeek. The “deepseek” row has a denominator of 4, not 1. No per-surface figure in this edition compares with a surface in another edition, and the four pooled assistants cannot be separated from the data published here.

• brands[].providerBalancedPct

  Provider-balanced visibility is neither computed nor displayed here. It would require every surface row to be a real surface, and one of the seven is a bucket we built; besides, with one answer per assistant, an average across providers is only a different weighting of the same ten answers. The rule that the two visibility metrics travel together (pooled + provider-balanced) applies to tier A editions, not to this one.

• brands[].visibilityCi

  No confidence interval is published, and therefore no statistical group column either. With one question and one answer per assistant there are neither clusters in which to measure an intraclass correlation nor a second run from which to estimate model variability. An interval on n = 10 would suggest that the number at its centre estimates a quantity in the market; it does not — it is how often ten models, asked once, picked the same name.

• brands[].avgPosition

  `avgPosition` is a pilot field: the mean of the positions in the slots where the brand appears. Below the 5-mention threshold it renders as “—”, which here means 27 of 29 rows; for the 19 brands with a single appearance, the “mean” is that one position. Only TEILOR (6 mentions) and Malvensky (5) clear the threshold. The values here are computed from the raw file, not from the published aggregate, which rounds to one decimal.

• metrics.mentionsPerAnswer

  Mentions per answer is exactly 5.00 (50 / 10), fixed by the “top 5” requirement. The metric is shown ahead of the ranking, as everywhere else — but here it is a constant of the instrument, so it is the one exception to the rule that market metrics may be compared across markets.

• metrics.silenceRatePct

  The silence rate is 0% by construction: all 10 answers name brands, because they were told to. The figure says nothing about the market and does not compare with the silence rate of a tier A edition, where the question allows an answer with no name in it.

• metrics.top1ConcentrationPct

  First-place concentration can be counted — five brands ever held position 1, and Malvensky holds it in 4 of the 10 answers — but it measures how often ten models agree on a leader, not market concentration. Each answer produces exactly one first place, by construction, so the ten first places are necessarily split between brands.

• sources

  There is no source table. The question explicitly asked for sources and some models supplied them, but they were not collected and coded systematically in July 2026. This is a declared absence, not a zero: `sources`, `denominators.sourceRate` and `counts.distinctCitedDomains` are null, and the sources section renders as absent with this reason on screen. With no sources there is neither a “tracked brand domain” column nor a “third-party sources only” filter.

• questions

  There is no per-question table. With one question, such a table would repeat the brand table row for row and make the same 50 slots look as if they had been counted twice. The question is reproduced verbatim in `market.universeRule` and in the DOI deposit; `method.questionsCsv` stays null because there is no questions CSV to reproduce.

• market.published

  Three distinct dates, never mixed: the measurement window is 13 July 2026 — a single day —, the study was published on 17 July 2026, and `editionId` “2026-07” follows the window, not the date on which the edition was transcribed into the index schema (8 August 2026, in `market.updated`).

Websem (2026). Romania's AI Visibility Index, edition 2026-07 — Luxury jewelry. Measurement window: 13 Jul 2026; 10 answers analysed. DOI: 10.5281/zenodo.21724399. CC BY 4.0. https://websem.ro/en/resources/aeo/ai-visibility-index/bijuterii-lux/2026-07