# Gr8 Index v1 — measurement dataset

The per-run measurement data behind every figure in the Gr8 Index v1, published
so that any figure can be **independently recomputed**.

Generated deterministically from the existing 1 October 2026 measurement log.
No new API calls, no new measurement — the same provenance as the v1.2 restated
tables.

## What this is, and what it is not

This is **counts and positions only**. It contains no raw answer text, no
verbatim AI prose, no citation URLs and no model identifiers. The complete
answer log is retained and available for independent verification on request;
it is not published here.

Each row records, for one brand in one AI answer, whether the brand was named,
where in the answer it first appeared, and whether that answer cited the
brand's own domain.

## Schema

| Column | Meaning |
|---|---|
| `category` | C1 agencies · C2 short-term insurance · C3 retail banking · C4 medical aid |
| `brand` | Canonical brand name, as frozen in the brand universe (Gate A) |
| `engine` | openai, anthropic, gemini, perplexity |
| `run` | Run number, 1–8. Each question was asked eight times per engine |
| `prompt_id` | Which question. Identifies the question, not its text; the verbatim question set is published in the Standard |
| `mentioned` | 1 if the brand was named in that answer, else 0 |
| `position` | Order of first mention within the answer, 1 = first. Empty when not mentioned |
| `cited` | 1 if that answer cited the brand's own domain as a source, else 0 |

Rows: **36,448**.

| Category | Brands | Answers | Rows |
|---|---|---|---|
| C1 | 28 | 544 | 15232 |
| C2 | 17 | 544 | 9248 |
| C3 | 12 | 544 | 6528 |
| C4 | 10 | 544 | 5440 |

## How to recompute a Gr8 Score

The unit of observation is one **(engine, run) cell**. Within a cell, over that
run's qualifying answers:

```
MR    = 100 * (answers naming the brand) / (qualifying answers)
SoV   = 100 * (sum of the brand's 1/position weights)
            / (sum of ALL tracked brands' 1/position weights)
CR    = 100 * (answers citing the brand's own domain) / (qualifying answers)

composite = 0.40*MR + 0.35*SoV + 0.25*CR

Gr8 Score = mean(composite) across all (engine, run) cells
```

Position weighting is harmonic: a brand named first in an answer carries weight
1, second 1/2, third 1/3, and so on. The SoV denominator is the sum of those
weights for **every tracked brand** in the same answers, which is why the
dataset carries a row per brand per answer rather than only the mentions.

Components are expressed as percentages and are **not** rescaled against the
field — a min-max rescale would make one brand's score move when a different
brand's did, which would break comparability between cycles.

## How to recompute the v1.2 intervals and ties

Keep the per-cell composites rather than only their mean.

- **Interval.** Engines are **fixed strata**, not a sample from a population of
  engines. Compute the variance within each engine, combine across engines, and
  the 95% interval covers exactly one thing: run-to-run variation of this panel,
  on this question set, on this date. All runs share one date, so every figure
  is a **same-day repeat** — the interval measures within-day jitter and says
  nothing about drift over time.
- **Ties.** Pairwise, never chained. Two brands are separated only when the 95%%
  interval of the **paired** difference — differenced cell by cell over the same
  (engine, run) cells — excludes zero. Pairing matters: both brands were
  measured on the same answers, so differencing inside the cell removes the
  shared answer-to-answer variation that says nothing about either brand.
  Indistinguishability is not transitive and is not asserted to be.
- **Direction.** A separation where every engine agrees on the direction reads
  "clearly ahead of". A separation that holds on the panel average while an
  engine dissents reads "ahead on panel mean only", and is not a settled
  ordering.

## Provenance

- Measurement date: 1 October 2026, logged out, region set to South Africa.
- Panel: four assistants, eight runs per question per assistant.
- Band: `High · API Panel · 4 engines × 8 runs · 1 day`.
- Method: the GR8ER2 AI Visibility Standard v1.2.
- Correction: CORR-002.