Methodology

Deduced from what players wrote

No player fills out a Buy-Again survey. The score is read out of the reviews people already wrote for a different reason, on the game’s own storefront page, in their own words. Two separate readings are taken from that same text and reported as two ledgers, never blended into one silent number:

• Explicit:reviewers who said outright that they’d buy it, play it, or recommend it again. Precise, but only the people who chose to say so.
• Inferred: what the rest of the review implies, even when the reviewer never stated an intention directly. Covers far more of the corpus, but is a deduction rather than a direct statement.

Both numbers are shown on every game’s page, next to the count of reviews each one came from. One is never presented as if it were the other.

Checked against real recommendations first

Before any score went live, the same reading method was checked against 1,473 reviews where the reviewer had also left a direct, separate recommend-or-not verdict on the platform. The method had to agree with those real verdicts at a level well above simply guessing the most common answer (90.5% accuracy against an 86.4% majority-class baseline) before it was trusted to read reviews where no such verdict exists. That check also surfaced a specific limitation, described below, rather than being quietly filed away.

Minimum evidence, always

A game with too few reviews does not get a score for a dimension it doesn’t have enough evidence for. Where that happens, the page says so plainly instead of showing a number. A handful of reviews also cannot swing a score to the extreme top or bottom of the scale; confidence has to be earned with volume, the same way it would from a person.

Every game in this edition is scored from at least 22 reviews, and most from 80. Where a game sits below the register’s usual depth, its own page says so.

Dimensions that didn’t measure consistently enough are not shown

The instrument reads for ten different qualities in a review, things like calm, social glue, emotional impact, price fairness, and more. Each one was tested for whether it gives the same answer when read from two separate, non-overlapping batches of reviews for the same game. Only three of the ten passed that test across this full register: calm, social glue, and emotional impact. Those are the only three shown here.

The other seven are not published, not softened, not shown with a caveat attached: they are left out of this edition entirely, because a measurement that can’t agree with itself isn’t a measurement yet. That withholding is deliberate, and it is the actual trust story: nothing appears on this site that hasn’t cleared the same evidence check as everything else.

A broken promise is not the same as the wrong buyer, but that reading is withheld here

Negative language in a review is not one thing. A reviewer describing a promise the game broke is a different signal from a reviewer who simply wasn’t the right audience for it, too hard, too slow, too grindy for their taste, and the instrument is built to read them separately so only the first kind would ever lower a score.

For this edition, the broken-promise reading is one of the dimensions that did not clear the split-half reliability bar described above, so it is withheld along with the other six: tracked, not published, and not part of the composite. The published Buy-Again Score for every game here is built only from the two buy-again ledgers, explicit and inferred, not from a promise-kept penalty.

When recent reviews disagree with the long-term record

A handful of games in this edition are flagged because their most recent reviews read differently from their all-time record, in either direction: sometimes recent reviews are unusually negative compared to the game’s history, and at least once in this edition, unusually positive. Each flagged game’s own page explains which way it points and why, rather than letting a temporary swing quietly move the number.

Frozen editions

This is the 2026-07-20 Edition, frozen on 20 Jul 2026. It does not get quietly edited after the fact. If the evidence changes enough to justify a new read, the result ships as a new, separately dated edition, and the old one is not rewritten to match it.

Scores are not for sale

No developer or publisher can pay to raise a score, and none can pay to have a low score hidden or left off the register. A score built from public reviews stays built from public reviews, whatever anyone offers.