Book market — visibility in AI answers · edition 2026-07

Across the 180 answers analysed, 247 mentions of the 21 tracked brands were counted — 1.37 brands per answer. 105 of the 180 answers name none: 58.3%, 95% confidence interval 48.0–68.0%.

In a market where the engine names 1.37 brands per answer on average, a visibility of 50% is harder to reach than in one with 4.34 — the case of the “Electronics and IT” market, measured over its own 87 answers. Density is a market metric, not a brand metric: it compares across markets precisely because it carries no name.

Concentration: the leading brand holds 15.0% of the edition's 247 mentions.

Each bar is one brand's distribution of mentions across positions, as a percentage of its own mentions — not of the answers. A brand named rarely but always first and one named often but always fifth have the same presence and not the same outcome.

The panel is 5 surfaces belonging to 3 providers, not 5 independent engines. The matrix footer shows how much of the sample each provider holds — the reason the columns cannot be averaged against one another, and the reason the ranking also carries the provider-balanced VEF column.

None of them contains a brand name. The set is published in full because it, not the ranking, is the reproducible part of the measurement: anyone who runs it again gets their own figures and can put them beside these.

422 distinct domains were cited in this edition; the table publishes the top 40. The list belongs strictly to this market: domains cited in another market count against another denominator and do not add up with these. The “third-party sources only” filter is on by default, and the column marks the tracked brands' own domains.

The two do not coincide. A brand can be named in the answer without its site being cited, and a site can be cited as a source without the brand being recommended. The gap between the bars is the difference between being talked about and being used.

The tables on this page are derived from these files, not transcribed. The licence is CC BY 4.0: they may be republished, with attribution.

• ↓ Brands (CSV)one row per brand: mentions, positions, citations, per-surface breakdown
• ↓ Questions (CSV)the 18 questions, verbatim, with the per-question mentions
• ↓ Cited sources (CSV)the cited domains and the number of answers citing each

Licence

CC BY 4.0

DOI

10.5281/zenodo.21736211

Mirrors

Zenodo — canonical data deposit, concept DOI 10.5281/zenodo.21736211, CC BY 4.0GitHub — the same three files, with history and verifiable checksumsKaggle — the three CSV files, with a column dictionaryHugging Face — dataset card and viewerdev.to — English republication of the analysis, canonical pointing at the study on websem.ro

Answer archive

None. The answer texts were not archived, so no hash can be published. The counts remain reproducible from the files above, but the answers themselves cannot be re-read. The absence is declared rather than hidden.

This page is a data sheet. The interpretation, the context and the implications live in the study-series episode, which uses exactly the same counts.

• Two runs, two weeks

  Enough to see the volatility — between the two runs Editura Corint quadrupled its mentions (2 → 8) and Editura Trei more than halved its own (12 → 5) — but not enough for a trend. Even these two moves clear the reporting threshold only if their intervals are disjoint, which at 180 answers they are not.

• 18 questions, not the whole purchase universe

  The panel covers the big territories — children, low budget, personal development, psychology, English, fantasy, cookery, the preparatory school year — but not exhaustively: textbooks, technical books and comics are missing. A brand that is strong in exactly what is missing looks weaker here than it is.

• 21 brands tracked, not the whole market

  Players outside the list — librex.ro, clb.ro, librariadelfin.ro, booknation.ro, printrecarti.ro — appear constantly in the citations, but their mentions are not counted. The 58.3% silence rate means “none of the 21 brands”, not “no shop at all”.

• Ten answers per question

  At the level of an individual question, the difference between 3 and 4 mentions cannot be told from noise. The question table is useful as a map of territories, not as a ranking: six of the 18 questions gathered at most three mentions in total, across all surfaces and both runs.

• Citations have a different denominator

  Cited-domain rates run against the 172 answers that contain citations, not the 180 analysed. A source percentage and a mention percentage on the same page cannot be subtracted from one another.

• Position, for two brands only

  The position breakdown was published for Libris (24 first places out of 37 appearances) and BookZone (3 out of 24). For the rest of the universe only the weighted sum exists, so “how high it appears” cannot be compared across the whole list.

• Origin and profile not recorded

  The queries were issued in Romanian, with no personalisation and no history, but where they actually originated was never recorded and never checked for consistency between runs. A logged-in user with a purchase history may get different answers.

• Mentions are not sales

  What is measured is a name's presence in an answer, not traffic, real market share or service quality. A brand invisible here can be the market leader; the overlap between the two hierarchies is the subject of study, not an assumption.

• Conflict of interest, declared

  BookZone is the edition's anchor brand and the only Websem client in the table; it appears with the same counts as anyone else, 4th of 21 by mentions and 5th by provider-balanced visibility. The study was funded by no analysed brand, and no brand saw it before publication.

There is no mechanical rule. The 21 brands are the competitor set configured in this edition's LLM Pulse project — online bookstores, marketplaces with a book section, and the largest publishers — chosen editorially, not by a threshold in a public ranking or an external register. The consequence is visible in the data: players outside the list turn up constantly in the answers (librex.ro, clb.ro, librariadelfin.ro, booknation.ro); we see them in the citations but do not count their mentions. The universe is therefore a declared choice, not a reproducible selection.

Universe version

carte-1.0

carte-1.0 · 2026-07-18

First version of the universe: 21 brands, the LLM Pulse project's competitor set at the opening of the measurement window. Editura Aramis enters with zero mentions and stays on the list — a tracked zero says something an absence does not.