The first benchmark measured to the Standard.

What the Gr8 Index is — and how it relates to the Score and the Standard

Three terms, cleanly separated, so nothing is conflated downstream.

Scope and engine panel (v1)

The four categories — South Africa

Four categories are measured in the first cycle, chosen for finite, recognisable brand universes, high stakes in an AI recommendation, and genuine buyer intent to ask an assistant. Each is measured identically.

The engine panel — four Core direct-API engines

v1 measures the four Core direct-API engines of the Standard, each queried through its official API with web search enabled where supported, each computed and reported separately (M-4), never blended silently.

Four engines with eight runs each clears the Standard's High confidence floor (≥3 engines and ≥8 runs). The panel, and any engine that fails or is excluded mid-cycle, is always named with the reason (R-3 graceful degradation): if Perplexity is unavailable, the Index runs and publishes on the remaining three — still High — and says so.

Which model, on each engine

Each engine is queried at the model tier a logged-out, everyday user is served by default — not the top-priced flagship — with web search enabled. The Index measures what an ordinary South African buyer actually sees, and an anonymous consumer is not served the most expensive model; choosing a representative tier is therefore a fidelity decision first and a cost decision second. The exact model identifier for every engine is logged per answer (M-8), published with the Index, and frozen for the cycle; changing a model is a version event (M-9), never silent.

The brand universe — engine-derived, by published rule

The brand set is the denominator of Share of Voice. Whoever picks it decides who can win. So we do not pick it: the engines reveal the field, a published rule decides inclusion, and the whole set — plus the brands that fell below the line — is disclosed.

The derivation, step by step

Entity resolution — the South African ambiguity problem

SA brand names collide across sectors, and a benchmark that miscounts them is wrong at the source. The resolution rules are fixed and published:

• Aliases are canonicalised: "FNB" → First National Bank; "Std Bank" → Standard Bank; "Absa", "Nedbank", "Capitec", "TymeBank" to their canonical entities. The canonical-name map is published with the Index.
• Multi-sector groups are resolved by category context. "Discovery" resolves to Discovery Bank in C3, Discovery Health in C4, and Discovery Insure in C2 — never pooled across categories. The same applies to Momentum, Old Mutual and Sanlam, which span life, health and short-term lines. A mention that cannot be resolved to the category's entity from its answer context is logged and adjudicated by a stated rule, and the adjudication log is kept.
• Company, not product: a product or plan name maps to its parent company entity.
• Global vs local: a global brand is included only if it genuinely serves South African buyers in that category and is named in SA-pinned answers.

The Prompt Corpus

The Standard's reference corpus is twenty prompts (9 Discovery / 5 Comparison / 3 Problem / 3 Branded) for a single-brand audit. A benchmark measures many brands at once, so the Index adapts the shape while preserving the intent proportions — and the adaptation is documented, as the Standard's Section 2 requires.

The Index corpus structure (the documented adaptation)

The 17 shared, brand-neutral prompts (9+5+3) are the comparable, non-self-referential measure of who AI recommends in the open field — these drive the league table. The 3 Branded prompts are instantiated for each brand (e.g. "what is {brand}") and feed only that brand's complete Gr8 Score, never the ranking, because a brand being asked about itself cannot be a fair comparator against the field. Per brand this is still 9/5/3/3 = twenty prompts; the intent proportions are preserved exactly (Standard §2). Prompts are weighted flat within each tier — no search-volume weighting, for the reason EB-038 makes plain: volume weighting re-hides the very choice the Corpus exists to expose. The two leader-anchored Comparison prompts are instantiated once per category from the top two discovered brands, run as shared inputs to every brand's Share of Voice, and printed in full in the published Index.

The four rules of a conformant Corpus (inherited, enforced)

C-1 Published — every prompt ships with the Index. C-2 Frozen per cycle — locked once the cycle begins. C-3 Versioned, never silently edited — superseded prompts marked and kept. C-4 Brand-neutral in Discovery and Problem — only Branded prompts name a brand.

The frozen v1 prompts for each category follow. Placeholders in braces are filled post-harvest and printed in the published Index.

Corpus scope note (v1): prompts are English-only. South African buyers in these categories query AI predominantly in English, but Afrikaans and other official languages carry real volume in insurance, banking and medical aid. English-only is a stated v1 limitation, logged in the blind-spot register (Section 6) and scheduled for a future multilingual expansion — never implied to be the whole picture.

The measurement protocol, instantiated

Every rule of the Standard's protocol maps to a concrete Index procedure. None is skipped.

Scoring and the league table

The three components, per brand, per engine

Mention Rate (MR) — share of qualifying shared answers in which the brand is genuinely named (after the Branded-Echo Guard). Share of Voice (SoV) — the brand's share of all tracked-brand attention, position-weighted: within each answer, distinct brands are ordered by first mention and weighted harmonically (position 1 = 1.0, 2 = 0.5, 3 ≈ 0.33, …); a brand's SoV is its summed weight over the summed weight of all tracked brands, averaged across answers, runs and engines. Citation Rate (CR) — share of answers citing the brand's own domain as a linked source, over all answers (M-10).

Components are computed per engine, then combined by an equal-weighted mean across the panel — engines are not weighted against each other, because there is no defensible basis to say one surface "counts more", and the per-engine figures are always shown so any reader may re-weight. The weighting of the three components (40/35/25) is the Standard's published editorial judgement, frozen per version, not a claim of optimality. Changing it is a version change, never silent.

Confidence band and interval — on every brand

With four engines and eight runs, every brand's Score qualifies for the High band. Each Score ships with a confidence interval computed from the observed run-level variance, sense-checked against the study's benchmark standard error of 0.062 (95% interval ≈ ±0.121) at eight runs EB-035. A brand whose interval overlaps the next brand's is reported as a statistical tie, never a false-precision rank — the Index shows the overlap rather than inventing a gap it cannot support.

What the Index publishes per category

A ranked league table of every included brand's Gr8 Score with its band and interval; the per-engine breakdown beneath each; the sentiment split; the no-retrieval rate per engine; and, where two brands tie within their intervals, the tie shown as such.

Blind-spot register — every known failure mode, and its control

A benchmark is only as trustworthy as its list of ways to be wrong. Each is named, with the control that contains it. This register is published with the Index.

Publication, cadence and governance

What ships with every published Index

• The ranked league table per category, each Score with its confidence band and interval, and ties shown as ties.
• The per-engine breakdown, the sentiment split, and the no-retrieval rate per engine.
• The full frozen Prompt Corpus, including the post-harvest leader instantiations (C-1).
• The frozen brand universe, the inclusion rule and thresholds, the below-threshold tail, and the canonical-name / adjudication notes.
• The engine panel, model IDs, run dates and conditions — and any engine lost or excluded, with the reason.
• The method version and a link to this document.

Cadence

Recommended: a quarterly public Index (stable enough to be newsworthy, frequent enough to show movement), with optional monthly internal runs for the ROI Lab. Each public cycle is a version; movement between cycles is reported only under M-9.

Home and governance

Proposed canonical home: gr8er2.com/index, a sibling to /standard, each Index cycle a dated, permanent entry. Governance follows the Standard's Section 8: the method is versioned, superseded elements are marked and kept, and a change to any definition, threshold, weight or panel is a version bump with a change log — never a silent edit. This document is the v1 method of record; it is itself under that rule.

Execution runbook & today's critical path

What has to happen, in order, to run the first cycle — and who does each part. The method above is complete; this is the operational path to a published number.

Dependencies to clear first

Run order

Data schema (minimum fields to log per answer)

category · prompt_id · intent · run_no · engine · model_id · timestamp · region_pin_method · region_pin_confirmed · retrieved_web(bool) · raw_answer · brands_named[] · brand_positions[] · citations[] · sentiment_per_brand · echo_guard_flag