Methodology
How FoeGlass turns AI answers into one comparable index.
What a single measurement is
A single measurement is one run: a specific phrase asked of a specific AI model from a specific geo, in a specific language, under a specific player profile (beginner, regular, high roller), with the collection date recorded. The model’s answer is stored verbatim. From one answer we extract one observation per named brand: its role (recommended / mentioned / warning), tone, position in the answer, and the sources it cites. Beyond that, FoeGlass analyzes overlaps in mentions across brands, geos, games and other factors. All data is collected manually by real people without using APIs, only from the described geos, account types and at the stated times. On top of that, the markup is done by two independent AI models that check the accuracy of the information after the human and after each other. The result is the most unbiased and transparent information we can produce.
The calendar
Data is collected and published in cycles; one cycle ≈ 60 days. Publishing fixes the snapshot: verbatim answers and observation verdicts are immutable. If the aggregation rules are extended (for example: languages, personas and so on), both visible snapshots are recomputed under the same new rules so that the N vs N−1 comparison stays honest, and the change itself is recorded in the changelog. The public view shows two snapshots: the current one (N) and the previous one (N−1); deeper history is kept but available only in reports.
How measurements become a score
Every observation is scored on four signals: the role the model assigned to the brand, the tone it used, how high the brand stands in the answer, and the quality of the sources the answer relies on. Adjacent measurements are not averaged blindly: panel cells carry different weights depending on geo, model and query intent; money intents weigh more than informational ones. A brand’s final score is the weighted average across all measured cells, from 0 to 100. The exact weights are part of the locked methodology and are not disclosed; the formula itself does not change within a cycle.
Cell states: unmeasured ≠ zero
Every panel cell is in exactly one of four states:
the brand was measured in this cell; it has a score.
the cell was measured, a market exists there, and the brand is not in it. This is an honest zero: the deposit goes to someone else.
the cell was measured, but there is no market in it (see §5). Such a cell is fully excluded from the calculation; it is neither a zero nor a penalty.
the cell has not been measured yet. It contributes nothing to the score and is never shown as "bad"; the absence of a measurement is not data.
The principle: "not measured" and "bad" are different things, and the interface never colours one as the other.
The witness rule
A red verdict requires proof that a market exists. A brand’s absence from an answer counts against it only when the same answer contains a witness, at least one tracked brand from the reference set, matched confidently and framed non-negatively: the model recommends it or mentions it neutrally. If the model names none of the tracked brands, or says the category is unavailable here, the run is classified as "no market" and leaves the denominator instead of being counted against anyone. A negative mention is not a witness, but for the mentioned brand itself it is real presence, with a negative sign. Verdicts are frozen at publish together with the brand dictionary version: later dictionary changes never rewrite published history.
Corrections and changelog
Verbatim answers and verdicts of published cycles are immutable. Every methodology change (a correction or an extension) is recorded here; when it affects aggregates, history is recomputed under the new rules symmetrically.
the index was extended to all measured languages (en+pt in current data). B shifts stayed within ±0.8, the top-5 ranking did not change; the main effect is that Brazil (pt) is visible in default slices for the first time.
Principles
The score cannot be bought. Paid work with brands provides depth of data and help with genuinely improving visibility; it never edits the index. Public data is untouchable: if we find an error, we fix it from the next cycle and record it in the changelog. All rights to the brands mentioned belong to their owners. FoeGlass is an analytics platform.