Tags describe the box, not the machine

Steam has about 450 community tags. They mix genre, mood, setting and business model with a handful of mechanics, and players vote on them. A tag list is popular consensus, not design analysis. The only peer-reviewed study of Steam's tag vocabulary concluded that mechanics are the largest gap in game metadata: the thing players talk about most and catalogues describe least.

So I built the machine view. Every game gets a design profile, with its premise, core loop, progression, failure and pacing, and a list of mechanics from a closed, maintained vocabulary. Each mechanic is either something you do or something that pushes back, and each is marked core, supporting or other for that game. Mechanics connect through eight kinds of link, and that wiring is what powers similarity.

The founding example was Hollow Knight. An early version of its profile listed Charm Equipping and Healing as the core mechanics. Both correct. Neither tells you what the game is. Melee combat, exploration and metroidvania progression sat in the "other" pile next to map annotation. The rule that came out of that now runs through the whole site: a game's core list must let a reader reconstruct what the game is about from the list alone. Correct is not the same as useful.

The idea, in four claims

1. A mechanic is a verb or a rule

   Twenty years of game studies produced at least 49 competing definitions of "game mechanic" and no experiment to decide between them. The two big camps are "mechanics are actions" and "mechanics are rules". SteamMaho keeps both and makes the split a label on every entry. Verbs are things the player does: Dash, Crafting, Pogo Attack, Lock Picking. Rules are always-on systems that push back: Death Penalty, Fog of War, Permadeath, Weather System. The vocabulary splits almost evenly between the two. Games really are about half what you do and half what pushes back.

2. Importance belongs to the pair, not the mechanic

   Dash is in thousands of games. In most of them it is transit. In a few, the game is substantially about dashing. So importance is a property of the game-mechanic pair, and every mechanic on a game page sits in one of three tiers. Core is the game's identity, the list you could reconstruct the game from. Supporting is the toolkit around that identity. Other is present but incidental. The median profiled game carries 10 mechanics, 3 of them core. Mechanic pages show games built around a mechanic separately from games that merely have it, a distinction no tag system can make.

   

3. Wiring carries identity

   A list of parts does not explain a machine. The wiring does. Mechanics connect through eight kinds of link: needs, works with, competes with, part of, unlocks, modifies, feeds, turns into. Six come from the literature, two were forced by annotation practice. The catalogue carries about 75,000 links. "Works with" is the most common and "turns into" the rarest, which is what the theory predicts: synergy is the everyday wire, true conversion is special. A link counts as confirmed only when more than one independent analysis pass finds it, and every game page marks each row as hand-verified, confirmed, or single-pass.

   

   

4. A core list must reconstruct the game

   Tiering has written rules, iterated against real games. The reader test: can you tell what the game is from the core list alone? The hook rule: a game's selling point must not be dropped, so 33 Immortals keeps its 33-player co-op in core. The genre-mate test: would this core list separate the game from its genre neighbours? One scene is not core. Aesthetics are not mechanics.

   Words also come in sizes. Jump King is about jumping. That is a small, precise word, and it is the whole game. Hollow Knight is about movement as progression, a bigger word with members inside it: dash, wall jump, double jump. The same vocabulary has to work at both altitudes, so it has umbrella mechanics with members, and a game's core list may sit at whichever altitude describes it.

What is on the site

Everything below is on the public site, no account needed.

• Game pages. Design profile, mechanics by tier, the wiring list and a one-click graph, similar games in two lenses, the story line, player-count history, price and follower charts, owner and revenue estimates with ranges, news, top reviews, and a "suggest a mechanic" button.
• Similar games, with a calibrated number. Ranked by shared mechanics, rare ones counting more, plus shared wiring. Two different games can never share everything, so the scale is anchored where true sequels actually land: 90%+ reads as same series or near-twin, 60%+ as close kin. A second lens matches store-page text and keeps its own number on purpose.

• Mechanic pages. 8,299 of them: definition, verb or rule, category, games built around it versus games that have it, related mechanics. The full vocabulary is a nightly CSV, free under CC BY 4.0.
• Discover. Pick several mechanics and see every game that has all of them, plus the most common pairs.
• Map of Steam. About 80,000 released games on a zoomable canvas. Distance is design similarity. Regions are named automatically from the mechanics their games share. A year time-machine, search, and pinned games with shareable links. The dataset is downloadable.

• Owners and revenue estimates. The core signal is a game's review count, corrected by multipliers calibrated against 247 publicly disclosed sales figures. Where a developer has disclosed sales, the disclosed number is shown with a badge instead of a model guess. Every estimate carries an 80% range, wider for games with few reviews. Revenue is owners times list price times a realized-price ratio, never the raw sticker price.
• Story structure, a second axis. Mechanics say what you do, not how the story is wired. Story-tagged games carry a separate label: linear, foldback, hub, modular, systemic, player-authored, or no authored story. Two readers labelled 200 games first and agreed with a kappa of 0.78. The automated labeller that covered 11,000 more matches them 84% of the time; two cheaper models were rejected at 69% and 71%. What the labels found: the Steam tag "Choices Matter" over-promises, since only 66% of games carrying it let choices change which events happen. "Multiple Endings" is honest at 88%. With genre, size, year and price held fixed, foldback stories run about 2.4 points higher in percent-positive reviews than linear ones.
• Research. A method page in six chapters, each anchored in a live exhibit. A working paper, "Atoms and Wiring: Describing Games by What They Let You Do". Fifty published articles. A public scorecard of every prediction the site makes.

Receipts: honesty as a feature

Any number that leaves the building gets graded in public, and the grading has to be tamper-evident. Every figure on the scorecard is a literal in a version-controlled file, never recomputed at request time. A recomputed receipt is not a receipt. As of 10 September 2026:

• Launch forecasts. A pre-registered 12-game backtest: 9 of 12 landed inside the published 80% range. The bar was 8. The three misses are printed, not hidden: two games exploded upward, one collapsed below its floor.
• Owners against disclosed sales. On 32 games with figures disclosed in the last 120 days: 17 within plus or minus 50%, 11 within 30%, median ratio 0.95x.
• An independent contrast. On an open, SteamSpy-derived panel, 61 games have a disclosed sales figure. SteamMaho's estimate was within 2x of the truth for 36 of them. The SteamSpy-derived figure managed 17.
• Story labels. Two-reader kappa 0.78, labeller 84%, and a table of how often each Steam tag agrees with the label.
• Open live predictions, logged before launch, graded at launch, never edited in between.
• Honest marks everywhere. Every wiring row says how it was found. Mechanic pages say "assigned by a language model, human-reviewed", never "curated by experts".

One feature was held back on purpose. A per-game freedom score, how much room a game's rules leave you to choose, computed from the wiring, passed every test designed to kill it. It is still not published as a ranking. Today it would measure how much analysis a game received rather than how much freedom it grants. The paper reports it as a negative result.

Scale

Read from the live database on 10 September 2026.

The median profiled game carries 10 mechanics, 3 of them core. The most common core mechanics across the catalogue: Exploration, Puzzle Solving, Turn-Based Combat System, Multiple Endings, Melee Combat.

What I took from it

• Correct is not the same as useful. A core list can be accurate and still tell the story backwards.
• Importance belongs to the pair. The same mechanic is the whole game in one title and transit in another.
• Flat lists lie about kinship. Wiring is what makes two games siblings.
• One inversion is a better proof than an average. Hollow Knight and Silksong, last by tags, first by wiring.
• Publish the misses. Three of twelve test games escaped their ranges, and all three are on the scorecard.
• Do not ship a measure that measures effort instead of the thing.
• The presentation of an estimate is the landing experience. The first developer to find the site judged it on the one number next to their own game.

Links

• steammaho.com the site
• How SteamMaho reads games the method page, six chapters
• Atoms and Wiring working paper, draft v0.5, August 2026
• Public scorecard every prediction, graded
• Map of Steam about 80,000 games by design similarity
• Hollow Knight a game page, with the wiring graph
• The vocabulary as CSV free to download, CC BY 4.0