Romania's AI Visibility Index

The first question is not who wins, but whether anyone is named at all. A market where the models stay silent in six answers out of ten is not run like one where they stay silent once in eighty-seven — and the difference is invisible to anyone reading a single market.

On Book market, 105 of the 180 analysed answers (58.3%) name none of the 21 tracked brands. On Electronics and IT, 1 of the 87 analysed answers (1.1%) names none of the 13 tracked brands.

Search by name or by domain. A brand present in two markets gets two cards, never one merged row: the same name scores differently in different categories, and the average of the two describes nothing.

Below is the protocol of the most recent edition — Electronics and IT, 2026-08. Every edition publishes its full protocol on its own page; the shared rules, the statistics and the dispute procedure live on the methodology page.

The full methodology — every question, each surface's denominators, the Wilson intervals, the rank groups, the thresholds, provider weighting, the dispute procedure, the errata and the method changelog — is a separate page: the index methodology.

What exactly does this index measure?

We measure how often AI surfaces name a tracked brand, across a fully published set of questions that contain no brand name. For each market we report in how many of the analysed answers each brand appears — for instance, across the 87 answers analysed on the electronics and IT market on 3 August 2026, eMAG was mentioned in 82. That is all we measure: mention frequency, in the declared window, on the published questions.

What does the index not measure?

We do not measure AI referral traffic: that would require access to every site's analytics, which we neither have nor ask for. We do not measure service quality, price, safety or competence. We do not measure conversions. And we do not produce a national per-brand score.

Why is there no national score per brand?

Because the same brand gets very different figures in different markets, and any way of combining them is a methodological choice rather than a measurement. eMAG is mentioned in 82 of the electronics edition's 87 answers and in 9 of the book edition's 180. Pooling the two gives 91 of 267 answers; averaging the two rates gives close to 50%; taking the market where it is a core player gives 94.3%. All three are defensible and they differ by some 60 percentage points. We publish none of them.

Why “5 surfaces from 3 providers” rather than “5 engines”?

Because three of the five surfaces are Google's — Gemini, AI Mode and AI Overviews — and they hold 51 of the electronics edition's 87 answers, 58.6% of the sample. “Consensus across five engines” would largely mean Google agreeing with itself. That is why every brand carries two figures: plain visibility, with a confidence interval, and a provider-balanced visibility, without one.

Why do you publish rank groups instead of places?

Because the data does not support as many places as there are brands in the table. Two brands share a group when their confidence intervals overlap. On the book market the top eight brands — from 37 down to 13 mentions out of 180 answers — form a single group: the hierarchy between first and eighth place does not exist in these data. The raw rank stays in the table, but the conclusions quote the groups.

Why are the confidence intervals so wide?

Because answers are not independent: each question produces up to five answers and they correlate with one another. We measured the intraclass correlation on the electronics edition's dense table and got 0.243, which gives a design effect of 1.93, rounded up conservatively to 2.0. Every interval is computed at half the answer count: 43.5 instead of 87, 90 instead of 180. The width is the information, not an excuse — it shows exactly how much of a ranking is noise.

My brand is not in the index. What does that mean?

Usually it means its market has not been measured yet. The index covers two comparable markets and one pilot edition today, so an absence says nothing about a brand's visibility in AI answers. If a brand is in a market's tracked universe and has no mentions, we say so explicitly, with the denominator in view: tracked across all of the edition's answers, with no mention recorded.

Can a brand buy, improve or delete its position?

No. There is no paid placement, premium profile or sponsored logo, and each market's universe is frozen before the run and published. Websem sells AI-engine optimisation services, including to brands in the index; all of them are marked in the tables. An edition's commercial anchor pays for a private run, on its own prompts, unpublished — never for anything connected to the public panel.

How often is the index updated?

The launch edition announces no cadence, because “updated monthly” with two markets would be a promise anyone could check on day one. A new market enters when it has a universe derived from a published rule. If we announce a date and miss it, we publish a delay note on the methodology page; we do not go quiet.

How do I cite the index, and how are the data licensed?

All data — tables, CSV and JSON — are licensed CC BY 4.0: usable, commercially included, with attribution. Cite the edition URL rather than the market URL: the market page always shows the current edition and will change, while the edition page stays as published. Recommended form: Romania's AI Visibility Index, Websem, the edition in question, the edition URL, the date accessed.