How it works

One log ruler, five games, the same puzzle for everyone each day. Guess how many Steam reviews each game has and try to beat a deliberately simple baseline.

The game

Five Steam games a day. Each round shows only the name, cover, price, release date, top 5 tags and a one-line description. You place your guess for the total review count (all languages, all review types) on the ruler. After five rounds you get a total out of 500, the number of rounds where you beat the baseline, a share text and a downloadable share image.

A new puzzle appears at 00:00 Beijing time (UTC+8). No account is needed; your history (scores, streak) is stored only in this browser.

Scoring

Each round is scored by the distance between your guess and the truth on the log ruler; guessing high and low are treated the same:

score = round(100 × max(0, 1 − |log10(guess / actual)| / 2))

Exact: 100. Off by 2×: 85. Off by 10×: 50. Off by 100× or more: 0. The red band on the ruler is that distance.

Boxes in the share text: 🟩 ≥ 80, 🟨 ≥ 50, 🟧 ≥ 20, ⬛ below 20. On the page, a vermilion dot next to a round's score means you beat the baseline in that round.

The pool and its distribution

The pool contains games released at least 60 days ago with ≥ 20 reviews, a cover, a description and at least 5 tags; adult-content games are excluded. The grey area under the ruler is the real distribution of that pool:

The five rounds come from the five review-count ranges separated by the red lines: 20–200, 200–2k, 2k–20k, 20k–200k and 200k+, in random order. Puzzles are determined by the date and do not repeat a game within 90 days.

After the reveal: peers and tag competition

Guessing a number should not be the end of it. After each reveal you get two more sets of figures, computed from the same pool:

Against its peers: peers are games with the same main tag released in the same year band; when fewer than 30 exist the group widens to "carries this tag (among its top 20) and same year band", then "carries this tag", and finally all games of that year band. The page says which definition was used. It draws the peers' review-count distribution, their median and the middle 50% range, and shows where this game ranks among them, how many times the median it is, and its price against the peer median price.

Tag competition: for the game's top 3 tags, the number of games in the pool, how many were released in the last year (60–365 days ago), the median review count, and the share with over a thousand reviews. Many games and many new releases mean a crowded tag; a high "≥1k share" means games in this tag break out more often. These answer the developer's questions: is this genre a red ocean or a blue one, and what does a typical game in it reach.

Peer data carries its own "as of" date, which can differ from the puzzle's frozen date. The figures state facts; they do not claim causes.

Forecast: unreleased games

Every Monday, five games due in 7–35 days. You predict their total review count 30 days after release. Change your mind until release; at the planned release time shown on the Steam store page the round locks; 30 days after release it is settled with the crawler's first review snapshot after that moment. Scoring is the same as the daily game.

Selection: among unreleased games outside the pool, those whose developer or publisher already has a released game come first (ranked by that game's review count in the pool, top 40); random newcomers fill the remaining places; five are then drawn with the week as the seed, preferring distinct main tags and release days. A game appears only once.

Settlement: the actual release time is taken from the store data after release. A round is void, and not scored, when the release slips by more than 60 days or the game is still unreleased 90 days after the announced date.

Ranking: per week, by the total over settled rounds; a round you skipped counts 0; the baseline model is ranked like a player. Before a round locks only the number of players is shown, not the distribution, to avoid herding; after the lock the distribution and median appear. The forecast baseline uses peers released 60–365 days ago with ≥ 20 reviews, so it runs high.

Identity and anti-manipulation: no account; the browser generates a random id and the leaderboard shows only the first 6 characters of its hash. One record per id per game (editable); at most 30 distinct ids per network address per week (the address is stored only as a salted hash); nothing is accepted after release. Rankings carry points and bragging rights only.

The baseline

Every round comes with a guess from a simple baseline (the dashed blue mark). It looks at three features only: the top-weighted tag, the release-year band (≤2014 / 2015–2018 / 2019–2021 / 2022–2023 / 2024+) and the price band (free / <$5 / $5–15 / $15–30 / ≥$30). Its prediction is the median review count of games with the same combination. When a combination has fewer than 30 games it falls back step by step: tag + year → tag → year + price → the site-wide median. The day's puzzle games are excluded from the calculation.

It is deliberately simple and never uses the review count itself as a feature. A proper prediction model arrives in a later phase.

Data source and freshness

Every field comes from public Steam endpoints (store item info and review summaries). A puzzle is frozen on the day it is generated: the answer is the total review count at that time and is not changed afterwards even as reviews keep coming in. Each puzzle shows its "data as of" date. Puzzles are currently generated through 2026-11-19.

Review counts are public Steam data. They are not sales figures and not our estimates.

Disclaimer

This site is not affiliated with Valve. Steam is a trademark of Valve Corporation. Cover images are hot-linked from Steam's official image URLs. The game awards points and bragging rights only; there is no real money or redeemable reward of any kind.