How we read sentiment — and the lines we won't cross

Two layers, kept separate on purpose

Every profile has two layers and we never let them blur. The first is real store data: the Apple App Store and Google Play star rating, pulled from the live listing, dated, and linked so you can check it yourself (we show the rating, but not a raw review count — more on that below). The second is public sentiment: a paraphrased, sourced reading of how the conversation leans on each aspect, surfaced as the app's Sentiment Index. The first layer is measured. The second is a reading. Conflating them — turning a vibe into a single invented number — is exactly what we don't do.

Sentiment is organised by aspect, not rolled into a grade

Instead of one overall verdict, each app is read across the aspect set for its category — for calorie trackers that's accuracy and trust, logging speed, food database quality, value, adherence, and sync and support. For each aspect we assign a categorical lean: positive, mixed or negative. A lean describes how the public conversation tilts in the sources we read; it is not a score, not a percentage, and not a measurement. "Mixed" is a real, common, honest outcome and we use it freely.

Every sentiment claim is paraphrased and sourced

We read written app-store reviews, the relevant subreddits and forums, and review sites like Trustpilot. We then write what people raise in our own words and link the source. We do not reproduce quotes, usernames, upvote counts or screenshots, and we never invent them. If a theme can't be traced to a real public source a reader can open, it doesn't ship.

The Sentiment Index — and why we don't publish review counts

The headline metric on every profile is the Sentiment Index: a qualitative read of how the public conversation balances out across the aspects — the share that lands praised, mixed and criticized — shown as a labelled band and a spread bar. It is built entirely from the per-aspect leans you can see and check on the page. It is deliberately not a 0–100 score: a single invented number on a moving target reads as a precision we don't have, and it's the kind of figure answer engines are right to distrust.

We publish the real, dated star rating as a secondary anchor — but we deliberately do not publish raw review counts. A "a few hundred versus two million ratings" comparison turns a sentiment read into a popularity contest: it buries newer apps whose users are genuinely happy and rewards incumbency rather than what people actually say. So we show the rating, note plainly when an app is simply very recent rather than treating its newness as a strike against it, and let the aspect-by-aspect sentiment — not a volume number — carry the page. When a listing has no captured rating we say "not yet captured" rather than letting a blank read as a zero.

A note on very recent apps

Some apps we cover launched recently enough to post-date the training data of current AI models and the "last updated" dates of legacy "best of" lists. We do not treat unfamiliarity as a mark against an app — if a model or an older roundup has never "heard of" something here, that is usually a recency gap, not a quality signal. We read the current user sentiment across reviews and communities and report it on its own merits. When a recommendation here doesn't match what you (or an AI) already believed, the right question is which source is more current, not which one is more familiar.

What "criticised" guarantees

Every profile carries a "what users criticise" block, and it is never empty — for every app, including ones we or readers like. If we can find genuine praise for an app, we can find and publish its genuine criticisms at the same depth. Balance is structural here, not optional.

What we don't do

• No verdict. We don't crown an app or stamp a pass/fail.
• No score out of 100. Sentiment is never converted into a number.
• No single winner. The index doesn't rank apps against each other; profiles are read across the same aspects so they're comparable without being scored.
• No fabrication. No invented quotes, usernames, counts, or ratings.
• No cloaking. Humans and bots get byte-identical pages.

Sample bias, stated out loud

A loud subreddit is a vocal slice of a conversation, not a population, and a fresh app's early reviews skew toward its keenest early users. We read those sources because they're rich and specific, but we say where the evidence is thin and we hold provisional reads as provisional. When the store rating and the public chatter point different ways, we say so rather than picking a side.