How the index is measured

01

Principles

02

What we measure: Core, Edge, Lens

Everything a customer sees belongs to one of three layers. Only one of them produces a score.

Edge never enters a score because a score a customer can tune is not a benchmark. What customers ask through Edge is still signal: aggregated Edge demand is reviewed at scheduled version reviews, where it can shape the next Core version. It never changes the current one.

A Core version change is a comparability boundary. It is dated and disclosed, and no trend, diff or alert crosses it.

03

The question set

A market's question set is built from real player questions with commercial and reputational intent, written in the market's own languages, weighted by real search demand, and categorised. Each question maps to one theme of the engine's fixed taxonomy (Trust & safety, Bonus, Payout speed, Product & app, Live casino and Sportsbook) or is held out as monitoring-only, which is asked and archived but never scored.

Demand weighting

A market pack supplies each question's monthly search volume, never a weight. The engine derives every weight with one formula, freezes it when the version is seeded, and refuses a pack that tries to supply its own:

Monitoring-only questions weigh zero everywhere. A market with no demand data weighs every question equally, and we say so rather than inventing precision. The per-market figures, including how many questions a set holds and its theme split, are published in each market's own pack disclosure.

Versioning

New questions mean a new version, never an edit in place. Stored answers and scores are append-only and preserved across versions. Trends, diffs and change alerts refuse to compare two scans of different versions: the comparison is declined, not fudged.

Questions never name a brand

No question text may contain any alias of any brand in the market's dictionary; the engine refuses to load a pack that breaks this. Because no question names a brand, one market scan measures every brand on identical input, and no brand can be handed a friendlier question.

Why the texts are unpublished

We publish the shape of the set and not the texts. A published question set is a gameable one: content could be written to the exact wording, and the benchmark would decay for everyone measured on it. This is the same logic that keeps Edge out of the score. The texts are fixed, versioned, and identical for every brand in the market.

04

The model panel

The active panel is 6 models: ChatGPT, Claude, Gemini, Perplexity, DeepSeek and Grok. Each is pinned to a dated release id, never a floating "latest" alias, so the instrument does not drift under the measurement. All 6 are scored.

The two columns matter for reading any result. A memory-only model answers from training-time knowledge: its answers reflect a slow consensus and move slowly, so a brand absent there has an awareness problem, not a news problem. A web-grounded run reflects what is retrievable today and can move week to week. Averaging the two without labels would be misleading, which is why every per-model view is labelled. On the citation track, ChatGPT, Claude, Gemini and Perplexity run web-grounded; DeepSeek and Grok are never web-grounded and contribute no citations.

Panel changes and retired models

The panel is capped at 6 models, and it changes only at a question-set version boundary, so no version ever mixes panels. A retired model stops being asked; its stored answers keep rendering in every historical scan. Retired so far: Llama. Every page reads the panel snapshot stored on the scan itself, never the current code list, so a panel change can never rewrite a past scan.

Request parameters are fixed: temperature 0.3, up to 500 tokens per answer. The question text is byte-identical on both tracks; web grounding is enabled through request parameters only.

05

Markets, licences, and why iGaming is different

A gambling brand does not exist globally. It exists per licence, per market, per domain: the same name can be a licensed operator in one market and an unlicensed one next door. A global visibility score averages those two facts into noise and hides the only thing that matters to a licensed operator: what AI models say inside the market where its licence applies. Mira Index measures nothing globally. Every number on every page belongs to exactly one market.

What a market pack contains

• A licensed-brand dictionary: every brand on the market's official licence register, with its aliases and domains, so detection reads answers the way a player writes brand names.
• Dated licence-status records: who is licensed, from when to when, approved by a named person.
• A demand-weighted question set in the market's own languages, versioned as described above.
• A watchlist of brands commonly named in that market without holding a licence there, plus an engine-level list of offshore brands licensed nowhere we measure.

Licence status is a dated lookup

When a scan is parsed, each detected brand resolves against the status record whose validity period contains that scan's start date. Licences lapse, renew, get revoked and get acquired. A product that silently rewrites its own past when a licence changes is not a measurement product. Here, a status change affects only scans from its effective date forward: every older scan keeps the status that was true on the day it ran.

Three mechanics back this up. Every scan records a hash of the exact dictionary it was parsed with. Brand and status records are append-only: a change closes the old record and adds a successor, and a brand is never deleted. And every recorded mention carries the licence status as resolved on that scan's date. Together they guarantee one thing: any stored result can always be read against the exact dictionary and statuses that produced it. They do not re-verify old answers, and they do not make the dictionary infallible; they make it accountable.

Unlicensed observation

When a model recommends a brand that holds no licence in the market, we record it, show it, and alert on it. It is never scored and never ranked. For a licensed operator this is recommendation share lost to operators outside the rules, on the exact questions players ask. For a regulator it is dated, quotable evidence of where AI answers send players. We record and report these observations; we claim nothing beyond that.

On the register itself, plainly: licence statuses are founder-curated. They are entered by hand, dated, and approved with a named approver, verified against the regulator's public register on a stated date. Automated register synchronisation is built but not yet live; when it runs, it will produce proposals for a human to approve, never automatic edits.

The engine is market-agnostic

The engine contains no market-specific logic. A market is data: a register, a dictionary, a question set. Adding a market adds no engine code, and every market runs the identical formula, cadence and rules published on this page.

Coverage today

06

The scoring formula

The whole formula fits in a few lines. With the stored answers, the market's dictionary and this section, an analyst can recompute any score we publish.

• A mention is a case-insensitive match of a brand's dictionary aliases in an answer, at word boundaries. We store the brand, its first-appearance position, and the sentence containing it.
• First-appearance position is counted among all detected brands, licensed and unlicensed alike. An unlicensed brand named first pushes every licensed brand down one position and costs it weight. A brand in neither list is invisible and pushes nothing.
• A sample is one asking of one question to one model; n is the number of successful samples in the cell. Failed calls never enter a denominator.
• The score ranges 0 to 100 and is reported to one decimal.

Every score carries a band: a 95% Wilson confidence interval on the appearance proportion, computed per cell and rolled up with the same demand weighting as the score. Fewer samples make the band wider. The band is always shown, never hidden.

The score deliberately excludes: unlicensed brands (recorded, never scored), monitoring-only questions, web-grounded runs, customer custom questions and watchlists, free lite scans, failed calls, and any model a scan's stored configuration marks as measured-but-not-scored.

We refuse to call a week-over-week change real unless the confidence bands do not overlap and the score moved at least 5 points. Below that bar the system stays silent, because a false alert costs more than a missed one.

07

Execution and repeatability

• Cadence. Every monitored market is scanned weekly, on the deployed schedule 0 6 * * 1 (Mondays 06:00 UTC). A sweep on */5 * * * * (every five minutes) advances running scans, and a scan completes the same morning.
• Sampling. Each question is asked 5 times per model on the score track, and 3 times per web-grounded model on the citation track.
• Failures and retries. A failing call is retried up to 3 times in-call with backoff, then archived as a failed attempt. A failed slot is retried on later sweeps, at most 5 attempts per slot, and only within 48 hours of the scan's start. Failed attempts are kept forever; readers use the latest successful attempt.
• Finalisation. A scan is terminal when every slot has a successful attempt, has exhausted its attempts, or the window has closed. It finalises with whatever succeeded.
• The reliability floor. A scan whose share of successful calls falls below 70% is marked partial: it is quarantined from comparisons, never a baseline for a change alert, never the source of a derived view, and a human is emailed the failure. Silence means success.
• Observed dates. The date shown next to any result is the date the answers were collected, never the date processing finished.
• Idempotency. Every call is keyed by scan, question, model, sample, grounding and attempt. An interrupted scan resumes exactly where it stopped and can never duplicate a call.
• Verbatim storage. Every raw answer is stored verbatim. Mentions, scores, citations and classifications are derived tables that can be rebuilt from the raw answers under a versioned parser; corrections are new rows, and old rows stay. Every rendered cell is a stored snapshot of what was actually returned, not a re-derivation.

08

Citations and sources

A citation is a URL a web-grounded response itself attaches to its answer. We capture the citation payload verbatim from the response, store it, and parse it into queryable rows under a versioned parser. We record what the model cited, not what we think it should have cited. Citations come from web-grounded runs only; memory-only models contribute none.

Every cited domain is classified into one of 6 source types: operator, affiliate, media, regulator, forum and other. Classification has three layers of authority. Deterministic rules come first: a brand-owned domain is an operator, known regulator and forum domains classify by list. A manual override by us is final until replaced. Everything the rules cannot decide goes through one model-assisted pass whose decisions are logged append-only and can be overridden. The mapping is not fully deterministic, and we do not pretend it is; it is rule-based where possible, reviewed where not, and every decision is on the record. Unclassified domains still count.

We refuse to rank sources on thin data: a domain with fewer than 4 scans of history or fewer than 30 recorded citations is listed as early data, without a rank.

For an operator, the source mix matters more than the score itself. The affiliates, review sites, regulators and forums that models retrieve from are where the answers are actually formed. The score tells you where you stand; the sources tell you which specific domains decided it.

09

What we do not measure

• AI surfaces outside the panel. Google AI Overviews, Google AI Mode and Microsoft Copilot are not measured. If it is not in the panel table above, we do not measure it.
• AI referral traffic, registrations and deposits. We measure visibility, not the click and not the conversion. The referral layer, the utm_source=chatgpt.com kind, lives in your own analytics, and that is where to look for it.
• Factual accuracy of what models say about a brand. Whether a model describes your licence status, payment options or availability correctly is not measured. Probe questions for this are drafted and not live; nothing on any dashboard today measures answer accuracy.
• Personalised and logged-in answers. Every call is a fresh API request with no user history, so we measure the neutral answer, not the answer a signed-in player with a chat history sees.
• Citations from memory-only models. DeepSeek and Grok are never web-grounded, so no citation data exists for them.
• Retrieval by location. All calls run from one infrastructure location. Grounded answers can differ with the searcher's location; we do not measure that variation.

This list exists because a tool that only publishes its strengths cannot be audited.

10

Known sources of variance

• Model updates under us. Providers change models. We pin dated release ids so the panel never drifts silently, and a deliberate panel change only happens at a dated version boundary. A provider can still adjust behaviour behind a pinned id; stored snapshots and the weekly cadence make such a shift visible as a dated break rather than an invisible one.
• Run-to-run variation. The same model gives different answers to the same question. We ask every question 5 times at a fixed temperature and publish the confidence band that variation produces.
• Prompt-order effects. Every question is an independent, single-turn call. No conversation state carries from one question to the next, so answers cannot be steered by what was asked before.
• Retrieval differences. Web-grounded answers depend on what search returns that day, and search varies by time and place. Grounded runs are sampled on a fixed weekly rhythm from a fixed location, and each scan stores exactly what was retrieved and cited.

Repetition, a fixed question set, stored snapshots and dated version boundaries limit each of these. None of them is eliminated, and a vendor claiming otherwise is describing a different universe.

11

The buyer's checklist

Questions worth asking any AI-visibility vendor, answered for this one.

12

Questions we get

13

Change log

• 15 Sep 2026First published.

Last updated 15 September 2026. Seen something on this page that does not match what we do? Email hello@miraindex.ai. We publish corrections in the change log above rather than silently editing this page.