Methodology/methodology
How it works, and how we check it
Written for a competent reader rather than for a data scientist. No number appears on this page that our benchmark did not reproduce on the day it was published, and every result carries the date it was measured.
What we compute
Avanti Scout reads the reviews a listing already has, and reports on the reviews. It does not verify credentials, does not contact a business before publishing, does not remove reviews and does not host them.
The analysis separates the reviews that deserve weight from the ones that do not, reads what remains for the things that matter in the category, checks whether the listing today is the listing the older reviews describe, and writes up only what those checks produced. A summary that strays beyond them is rejected and regenerated, never published.
How the Scout Rating relates to the platform rating
The platform’s rating counts every review equally. The Scout Rating is the same reviews with the set-aside reviews removed and the recent window weighted. The report always shows both and the difference between them, because the gap is itself a finding. Every set-aside review stays visible with the reason it was set aside; nothing is silently deleted.
How confidence is calculated, and what each word means
Confidence is computed by fixed rules from the evidence itself. It is never chosen by a language model. These four words are the only ones the product uses, they mean the same thing on every surface, and they are defined here rather than left to the reader.
- High confidence
- A large surviving sample, signals that agree, and most of the evidence inside the current window.
- Moderate confidence
- Enough evidence to form a view, with either a thinner sample or some disagreement between signals.
- Limited evidence
- A view is offered and is explicitly soft. Either the sample is small or the signals point different ways.
- Not enough evidence
- Not enough evidence to form a view. No rating is shown at all, and the counts that led to that are. This is a designed result, not an error and not a failure to load.
The fourth is the one that matters. A tool that always produces a number is a tool that produces numbers rather than findings.
What we never say
We describe bodies of reviews, never businesses and never people. We say that reviews show patterns consistent with incentivised or coordinated posting, with the count and the named signal. We do not say that a review is invented, that reviews were purchased, or that anybody acted dishonestly. Where a pattern is one the Competition and Markets Authority’s guidance addresses, we say so and cite the guidance; whether an offence occurred is the CMA’s determination and not ours. These rules bind every piece of copy in the product, including notification text, and this website.
| Never | Always |
|---|---|
| This business bought reviews | 34 reviews on this listing show patterns consistent with incentivised or coordinated posting |
| Invented review | Flagged: burst timing and template language |
| Dishonest trader | Reviews mention a final price above the quote in 9 of 41 cases |
Where the reviews come from
Every review is reached lawfully: through a platform’s official or licensed access where one exists, or through a licensed provider that carries the compliance obligation contractually. Where neither exists, we do not cover the platform. We never collect Amazon data ourselves, in any form, and this is permanent: Amazon coverage comes only through a licensed third-party provider.
What the browser extension shows you is analysed against the page you are already looking at, and is not collected by us.
The benchmark set
Every version of the analysis is measured against a versioned benchmark of real UK listings across hotels, restaurants, holidays, products and local services: documented public cases, listings the platforms themselves flagged, synthetic sets with manipulation planted at known rates, known-good listings, and hard cases where the correct answer is a confidence band rather than a flag. Results on our own synthetic injections are reported separately and never combined into an overall figure. The set is versioned, extended quarterly, and kept strictly separate from the development of the analysis it measures.
The standard every version must clear
No version of the analysis serves a report until it passes a fixed held-out test. To pass, a version must beat the previous version on detection, hold false positives on known-good listings under a fixed ceiling, call hard cases correctly at a rate we will not lower, and keep ratings stable on unchanged review sets. Fairness is part of the same gate: no category, no business size, and no writing in English as an additional language may be flagged materially more often than the rest. Language fairness is measured on real review text, by features of the text, never by any assumption about who wrote it. A detector that flags small businesses or non-native writers more often is not detecting manipulation, it is detecting unfamiliarity, and we would rather hold a release than ship that.
Affiliate networks we are joined to
None. Avanti Scout is joined to no affiliate network today, so no report carries a monetised outbound link and no page on this site earns a commission. When that changes, every network is named here before the first link goes out.
What we may claim, and when
No accuracy figure appears on this website, in a store listing, in a press release or in a sales conversation unless the published benchmark reproduces it on the day it is made. Marketing copy is bound by exactly the rule product copy is bound by.