Methodology & KPIs

Every number on this site is computed by fixed rules from public customer reviews. This page discloses those rules completely: the formulas, thresholds, windows and exclusions. How we prove the inputs were not tampered with is documented on the Trust & Transparency page.

1. What counts as a review

• We collect publicly available customer reviews from review platforms and app stores, per brand and market, and normalize them into one dataset.
• Canonical reviews only. A review counts once: duplicates (the same review found via more than one route) are detected and excluded, and reviews we have manually excluded for quality reasons are excluded from every count, average, quote pool and answer.
• Coverage honesty. Some platforms expose only part of their review history publicly (for example, roughly the most recent 200 reviews, or one storefront's most recent 500). We collect what is publicly reachable and disclose coverage rather than extrapolating.

2. Time windows

• Analysis window: 24 months. Aggregates, quotes, evidence and FAQ answers use reviews from the last 24 months, so pages describe the current company, not its history.
• Headline rating window: 12 months where a page shows a "last 12 months" figure.
• Trend charts are not windowed. Monthly trend series are the historical record and keep all months.
• Award window: the 12 months ending 31 July of the award year. Every award year is a closed, immutable period.

3. Star ratings (recency-weighted)

Displayed average ratings weight recent reviews more strongly. Each review's weight decays by 10% per month of age:

weight = 0.9 ^ months_old

average rating = Σ(rating × weight) / Σ(weight)

A two-year-old review still counts, but at about 8% of the weight of a review written this month. Review counts are never weighted: "7,330 reviews" always means 7,330 actual reviews.

4. Sentiment and themes

• Each review's text is analyzed by a language model and assigned themes from a fixed taxonomy (quality, value for money, customer service, delivery, billing and so on) with a sentiment per theme mention.
• Sentiment comes from what reviewers wrote, never from star ratings. A 4-star review complaining about billing counts as a negative billing mention. We do not convert stars into sentiment, because that would fabricate opinions nobody expressed.
• Positive share for a theme = Σ(weight of positive mentions) / Σ(weight of all mentions), using the same 0.9-per-month weights as ratings.
• Theme sentiment covers the analyzed portion of the corpus. Where analysis coverage matters for a comparison, the page states the coverage explicitly.

5. Quotes

• Every quote is a verbatim excerpt of at most 200 characters, machine-verified to be an exact substring of the stored review before it may be published.
• Every quote carries a link to the original review plus its source, date and rating.
• Translated quotes are marked as translated, and the verbatim original is preserved and shown.
• Quotes are never edited, paraphrased or "cleaned up". Ellipses only mark shortening.

6. FAQ answers

FAQ answers are generated strictly from review evidence and must cite verbatim quotes from the reviews that support them. When the reviews do not answer a question, we say so or do not publish the question, instead of inventing an answer.

7. Comparison pages

• Brands are compared per theme on recency-weighted positive share, not raw counts, so a large brand cannot win simply by having more reviews.
• A theme needs at least 5 classified mentions per brand; below that the page shows "insufficient data" instead of naming a winner.
• A brand enters a product-category comparison only with at least 25 in-window reviews (15 of them analyzed) mentioning that product.
• A category page is only published with at least 3 eligible brands. No two-horse "best of" pages.
• Each comparison page carries its own methodology note with these numbers.

8. Awards

All awards are computed from the review data under these gates. There is no entry fee, no nomination and no way to purchase an award.

• Window: the 12 months ending 31 July of the award year.
• Scoreboard listing: a provider appears from 5 reviews or mentions.
• Field floor: a category needs at least 25 in-window reviews per contender and at least 3 contenders in the relevant market, or no award exists.
• Overall winner: at least 50 reviews and a lead of at least 0.2 stars over the runner-up. A statistically thin lead earns nothing.
• Theme winner: at least 25 classified mentions and a lead of at least 5 percentage points in positive share.
• Main award: a composite across theme dimensions, requiring floor-clearing presence in at least 3 dimensions and a composite lead of at least 0.03.
• Retraction over embellishment. When we conclude a field's data is too thin to be credible, we retract or withhold the award publicly rather than publish a weak one. Locked award years are immutable records.

9. Dates and updates

• A page's "last updated" date (and its sitemap and HTTP equivalents) changes only when the page's content signature actually changes. Rotating quote selections alone do not count as an update.
• The first-published date of a page is set once and never moves.

10. Corrections

Publishing is versioned: every change produces a new dataset version, and each version is cryptographically committed (see Trust & Transparency). Corrections therefore appear as new versions with new dates, never as silent edits. To report an error: hello@citatio.ai.