Open data: the explainable game similarity dataset
Most similarity data says two games are alike and stops. This dataset says why, what is different, and who each one suits — in two languages, with the official store pages each claim was checked against. It is small on purpose: every row was read by a person against current Steam records before release.
Last updated 2026-09-02
What is in it
Version 0.1.0, snapshot of 2 September 2026: 50 unique pairs covering 82 PC games. Each pair is one row with a stable relation ID, both games’ IDs and Steam App IDs, three bilingual text fields (why similar, key differences, best for), a confidence label, and the source IDs, URLs and capture dates behind it.
The text fields are editorial judgments made inspectable, not player-behaviour ground truth. Official store pages validate identity and described features; they do not validate which game is better, and the data does not claim to.
Files
| File | What it is |
|---|---|
| gvg-explainable-game-similarity-v0.1.csv | 50 reviewed pairs, one per row: both games with stable IDs and Steam App IDs, why they are similar, what differs, who each suits (English and Chinese), confidence, source IDs and URLs, validation date |
| sources.csv | 82 official Steam store records the rows cite, with capture dates |
| DATA-DICTIONARY.md | Every field, defined |
| METHODOLOGY.md | How pairs were selected from the 743-row internal pool, and what was checked |
| validation-summary.json | The QA numbers: 50 rows checked, 82 App IDs resolved, 0 mismatches |
| LICENSE.md · CHANGELOG.md · CORRECTIONS.md · SHA256SUMS.json | Licence, version history, how to report an error, file hashes |
| Everything as one zip | The whole release, about 31 KB |
How it was made
The internal pool holds 743 directional recommendations from the site’s reviewed “games like” pages. A candidate had to be high-confidence, cite at least two resolvable official-store sources covering both games’ Steam App IDs, agree on slug, title and App ID, have no reverse-direction duplicate, and carry no dynamic numbers or promotional wording. That left 112 unique pairs; the first 50 by selection score were re-read and rewritten conservatively against the current Steam records, and all 100 endpoints matched.
The full account is in METHODOLOGY.md. Later versions will extend the same pool in reviewed batches; a version is never edited in place.
Licence and citation
The GVG-authored structure and annotations are released under Creative Commons Attribution 4.0. Use it, build on it, redistribute it; credit Game V Game. Game titles and store identifiers belong to their owners and are included only to identify the games.
Suggested citation: Miles (2026). GVG Explainable Game Similarity Dataset, version 0.1.0. Game V Game. https://doi.org/10.5281/zenodo.22250985 — the DOI is a permanent Zenodo archive of this exact release; the source repository is on GitHub and a mirror on Hugging Face.
Corrections
Found a wrong identity, a feature that is no longer true, or a translation that misses? Write in with the relation_id and a link to the evidence. Fixes ship in the next version and are listed in CHANGELOG.md with credit if you want it.
The pairs in the data are the same ones the site’s “games like” pages and comparisons are built from, so a correction here fixes the site too.