How IndexFair handles reviews

This page sets out what we publish, where the data comes from, how we screen for fake reviews, and how to dispute a rating.

IndexFair publishes a single composite score per brand per market, derived from a published methodology. Each score is accompanied by per-aspect breakdowns, the methodology version used, and the count of review signals fed into the calculation. We do not publish individual reviews or short user-review excerpts verbatim. Public aspect rows use aggregate counts and IndexFair-authored analysis; the underlying review text remains outside the public response, as described in our methodology.

Review signals are aggregated from a small set of public sources. We do not host a review form on indexfair.com and we have never accepted a review submitted directly by a consumer.

For crypto exchanges and wallets the review signal is drawn from app-store and consumer-review platforms; the register column reflects FCA crypto-asset (MLR) registration for exchanges. Consumer crypto wallets hold no custody and are register-less — their trust signal is a multi-source attestation consensus, not a regulator licence.

IndexFair has no affiliate relationship with any VPN provider and takes no payment for placement or score. VPN scores are derived from independent audits and aggregated user reviews. A rated provider may purchase independent analytics of its own performance — that is data, not score movement, and any such relationship is published by name on our subscriber register.

For each source we record the aggregate rating, the number of reviews contributing to it, and the timestamp of last ingestion. The source categories we read, and why some are weighted higher, are described at /sources. We deliberately do not publish the exhaustive per-source list or its weights, so the input mix cannot be gamed. We do not retain individual review text after extraction — only short aspect-level audit excerpts of at most 200 characters, which are never published, and metadata sufficient to reproduce the aggregate.

A deterministic Layer-1 filter runs over every review before it enters the corpus. A Layer-2 model-assisted screen then reads the review’s prose for authenticity. Both layers only decide whether a review is trustworthy enough to count toward a score — they can exclude evidence but never adjust the score itself — and the count of excluded reviews is shown on each brand page. We do not publish the exact thresholds or prompts.

Layer-1 verdicts are tracked in the brand record as reviews.is_likely_fake + fake_reason. Reviewers are never named. We do read public forums — they carry unresolved complaints that never reach a review platform — but we exclude anonymous forums that expose no account age or posting history, because that is what makes a coordinated campaign undetectable. The sources we exclude, and why, are listed at /sources.

The review-signal extraction layer and the Layer-2 authenticity screen both run on DeepSeek (deepseek-v4-flash). Anthropic (Claude Haiku 4.5) is the certified extraction fallback; a candidate model must pass a ≤5 percentage-point recall benchmark against a labelled reference corpus before it can replace the incumbent. The Layer-2 screen is a one-way valve: it can only exclude a review from contributing to a score, never raise it. Both provider roles and the merge-blocker are documented in the published methodology.

Defensive commitments. If a behaviour you'd expect from a review-handling platform is not listed here, assume we do not do it.

• ✗We do not commission, request, or solicit reviews from anyone.
• ✗We do not incentivise reviewers — no payment, gifts, points, or any other consideration.
• ✗We do not accept payment from any rated company for placement, ranking, or a change to its score.
• ✗We do not publish a statement about a company that is not bound to the specific record supporting it — every figure on a brand page traces to a named field in that brand’s data.
• ✗We do not store verbatim review text. Short aspect-level audit excerpts of at most 200 characters are the only text that persists after extraction, and none of them is ever published.

Operators, members of the public, and regulators may dispute a rating or request a correction via our contact form using the «Dispute a factual record» category. Substantive disputes — those concerning a specific brand fact, a regulatory status claim, or a published score — are acknowledged within 14 calendar days and resolved as quickly as the underlying evidence allows. Resolutions that result in a score change, a published correction, or a methodology clarification are logged in a public-disclosure register that surfaces here on this page once the volume crosses one entry; the register is admin-curated, not automated.

Disputes that turn on the methodology itself — e.g. the choice of a weight or a half-life — are routed to the public methodology changelog at /changelog; no methodology change is made without a versioned entry.

Substantive changes to this policy are versioned alongside the methodology in the public changelog. Minor editorial corrections are made in place without a version bump.

For the score-computation mechanics underlying this policy, see /methodology. For the upstream-source list, see /sources.