The methodology of the AI Visibility Index

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.

• brands[].mentions

  A brand's visibility does not cross the market border. eMAG is named in 82 of the 87 answers here and in 5.0% of the book edition's answers; the two figures answer different questions, put to different surfaces, over different denominators. A brand present in two editions has two separate measurements, not a national score: they are not summed, not averaged and not placed in one ranking.

• brands[].weighted

  Weighted visibility is an unbounded sum that grows with the number of answers — 87 here, 180 in the book edition. Compared across markets it measures sample size, not performance. It enters a comparison only as positionScore = weighted / 87, and only inside this edition.

• surfaces[ai_overview].answers

  The AI Overviews denominator is 15, not 18: Google did not trigger on three questions. Using 18 systematically underestimates every brand on this surface — evoMAG's rate moves from 33.3% to 40.0%. Any surface-level comparison with another edition must use that surface's own denominator, not 87 and not 18.

• method.runs

  One complete run, on 3 August 2026. There is no earlier edition of this market, so there is no delta, no trend, and the movement-reporting rule (disjoint intervals plus a mention change of at least 35%) has nothing to apply to. Differences against a future measurement will include platform volatility, which cannot be separated out here.

• brands[].bySurface

  The order differs between the two visibility metrics, and the difference must be read rather than hidden. On the pooled figure Flanco (48 mentions out of 87) sits ahead of evoMAG (47); on the provider-balanced figure evoMAG moves ahead of Flanco, because three of the five surfaces are Google's and evoMAG does better on ChatGPT (11 of 18) and Perplexity (11 of 18) than Flanco (9 and 7). The same inversion appears in the tail, between Amazon (1 mention) and iStyle (2).

• brands[].positions

  The position distribution was recorded only for the six brands with the most mentions. For the other seven it is missing, and the position score is not computed from weighted visibility alone: five of them are below the 5-mention threshold for per-mention metrics anyway.

• sources.items

  The source table publishes 60 of the 249 distinct cited domains; the tail below 2 answers is not published. The answers-citing column is not a sum of answers: one answer usually cites several domains, so the column does not add up to 87 and does not compare against another edition's distinct-domain count.

• provenance.rawArchive

  The raw texts of the 87 answers were not archived, so there is no hash certifying them. The counts are reproducible from the three public CSVs, but the study's qualitative claims — who said what, in what wording — cannot be re-verified at source by a third party.

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.

• denominators.sourceRate

  Source rates run against 172, not 180. 172 is the number of answers that contain citations; the other 8 were generated without links. The arithmetic pins the denominator: carturesti.ro, cited in 46 answers, is published at 26.74%, and 46/172 = 26.74% exactly, while 46/180 would give 25.56%. One working copy of the source table still carries a header declaring “180 raspunsuri” above a column named only as a share of answers; the published files — the CSV on the site and the Zenodo deposit — name the column `pct_din_raspunsuri_cu_citari` and declare 172. The source was not rewritten. So that it can sit beside an edition that reports against all answers, the same counts are recomputed over 180 in `ratePctRebased`, with the published rate kept alongside.

• surfaces[].mentions

  The five per-surface totals — ChatGPT 33, Gemini 78, Perplexity 30, AI Mode 56, AI Overviews 50 — are not published as counts. The source publishes rates per brand and per surface, as percentages of that surface's 36 answers. The counts invert exactly, round(rate × 36 / 100), and the inversion closes twice over: every brand row recovers its own total, and the five columns sum to 247. The deposit's data dictionary publishes precisely this check, with these same five figures. The brand rows nonetheless keep the source's shape — `ratePct` published, `mentions` null — so that nothing the source did not count passes as a count.

• method.queriesIssued

  Not recorded. The edition declares 180 analysed answers and 36 per surface, but does not separate queries issued from answers returned. Episode 3 does make the distinction — 90 queries, 87 analysed answers — so surface trigger rates do not compare between the two editions.

• method.deff

  deff = 2.0 is transferred, not measured on these data; the full reason is in `method.deffNote`. The practical consequence: the Wilson intervals here are as wide as episode 3's, but their width is not justified by an intraclass correlation measured on the book market.

• questions[].isDense

  The per-question table contains only the rows with at least one mention. A brand missing from a question is not a confirmed zero but an unrecorded one — hence `isDense` is false on all 18 questions, `answers` is null, and no per-question rate is computed from this file. The same gap is what makes the ICC unmeasurable.

• method.runs

  Two runs seven days apart, pooled into a single denominator of 180. A brand can be counted twice for the same question, once per run, so `mentions` counts answers, not questions covered. Episode 3 has a single run: comparing rates directly between the two editions sets a repeated panel beside a snapshot.

• provenance.rawArchive

  There is none. The raw answers were not archived, so nobody can rebuild the counts from the text; only the internal consistencies between the three CSV files can be checked, and those all close. The absence is declared, not hidden.

• brands[].positions

  The position breakdown was published for two brands, Libris and BookZone. For the other 19 only the weighted sum exists, so `firstPlaceSharePct` can be computed for those two alone. The position score, which derives from the weighted sum, stays available for every brand above the mentions threshold.

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`).