Steam Market Research 7 archive
Frozen research artifacts, September 7, 2026. Large Parquet datasets, raw API pages and the full announcement inventory remain in the local research directory.
# Catalog expansion and advertised variety Offline study of the September 5, 2026 public Steam catalog. Uses paid released games outside content descriptors 3/4, with current tags and descriptions, recorded dates and the previously verified KOF XV date correction. Main years 2019–2025; January–August 2025 and 2026 compared separately. 81,619 rows enter the union of these periods. Neither this metadata filter nor app IDs constitute manual quality screening or title/edition deduplication. The question is about breadth and concentration in current storefront positioning attached to release cohorts. Historical tags, real mechanical diversity, originality, influence, actual copying, and gameplay quality are not recovered from these data. Newer named tags can make old activities appear new. ## Measurements - All top-20 tags; a fixed transparent set of 181 activity/genre tags; and each game's highest-ranked tag in that set. Each game contributes one total vote spread across its retained tags. The primary-label representation gives it exactly one vote, including an unclassified label when necessary. - Shannon effective diversity is exp(entropy): the number of equally common labels giving the same entropy, not a count of true genres. Simpson diversity and top-ten shares are saved as additional concentration views. - Activity pairs use at most the first five ranked activity tags. Counts of observed pairs are supplemented by exact hypergeometric expected richness in a uniformly drawn 3,000-game sample, including games producing no pairs. A second measure counts expected pairs occurring at least five times. - Coverage sensitivities: exactly twenty returned tags; exactly the first three activity labels among games with at least three; at least 100 reviews; and a fixed broad ten-tag view. Those use common 1,000-game rarefaction, kept separate from the headline 3,000-game comparison. - Neighbor comparisons sample 3,000 games within each of 2019, 2021, 2023 and 2025. Tag overlap is Jaccard similarity; the eligible set requires at least five non-administrative tags. Lexical overlap uses TF-IDF words and bigrams in short pitches, screened for at least fifteen Latin words, English-oriented function-word presence and predominantly ASCII text. This text subset does not represent every language in Steam. A common current vocabulary/IDF is fitted across 2019–2025, not as a historical prediction exercise. - Nearest neighbors exclude self, then same normalized developer-credit string, then both same credit and publisher. Names do not prove different teams. Alternate editions can still appear. Two 2,000-game random replicates and a 3,000-game one-per-credit sample are retained. All comparisons within a configuration use equal sample sizes across years. The selected highest-overlap examples and random neighbor examples are saved. They include edition variants and formulaic puzzle descriptions. Neither similar text nor similar tags establishes that two games play identically. No inference of AI authorship or game lineage is made. ## Reproduction Use `../.venv/bin/python` from this directory: `study.py`, `neighbors.py`, `sensitivity.py`, `charts.py`, `validate.py`. All scripts are offline. Sources remain unchanged. The root network collection policy remains paused. `findings.md` contains the interpretation, also delivered in chat; CSV tables and frozen neighbor memberships support inspection.
# Catalog growth versus advertised variety The recent catalog grew much faster than its comparable variety of advertised activity labels. There is modest broadening and some redistribution, but no evidence here of either a doubling of creative variety or a general collapse into indistinguishable storefront pitches. The analysis uses current September 5, 2026 tags and descriptions attached to release cohorts. It cannot reconstruct historical genre vocabularies, actual mechanical differences, influence, originality, or how games play. App IDs can also distinguish editions of the same work. These are measurements of public storefront positioning within the documented surviving catalog. ## Scale versus variety For paid games outside explicit descriptors 3/4: | Measurement | 2021 | 2025 | Change | |---|---:|---:|---:| | Released game apps | 7,959 | 15,674 | +96.9% | | Distinct pairs among first five activity tags | 5,549 | 7,074 | +27.5% | | Expected distinct pairs in equal 3,000-game samples | 3,812 | 3,980 | +4.4% | | Expected pairs represented at least five times in those samples | 999 | 1,068 | +6.9% | | Share assigned to the ten most common primary activity labels | 59.7% | 56.8% | -2.9 percentage points | A larger cohort mechanically contains more rare combinations. The equal-size comparison removes that part of the increase. The result is not zero change: labels are somewhat less concentrated, and repeated combinations become a little more varied. But the increase is much smaller than the near-doubling of supply. Shannon effective diversity across the fixed activity-tag set rises from 72.6 to 79.0, and across primary activity labels from 40.4 to 45.5. These are concentration measures expressed as equivalent equally common labels, not counts of real genres. The broader all-tag effective diversity is nearly unchanged, 156.5 to 158.0. The same-calendar comparison for January–August shows 9,922 eligible releases in 2025 versus 14,125 in 2026 (+42.4%), while expected pair richness in 3,000 games rises from 3,946 to 4,094 (+3.7%). Again, the main movement is scale, with a smaller increase in comparable advertised breadth. ## The 2019 measurement problem In the 2019 cohort, the median game has nine returned tags and 35.2% have at most five. In 2021 the median is twenty and only 1.8% have at most five; in 2025 those figures are twenty and 0.4%. Treating those records as equally rich descriptions would badly exaggerate the apparent increase in variety. Restricting to games with exactly twenty returned tags reduces the apparent 2019–2025 shift substantially. Expected activity pairs in a common 1,000-game sample rise from 2,270 to 2,641 (+16.3%), compared with 1,397 to 2,379 (+70.3%) without the coverage restriction. The exact-twenty subset is selected, not a perfect correction; it demonstrates sensitivity to metadata completeness. The 2021–2025 direction survives the other checks, at modest magnitudes. With exactly twenty tags, equal-size pair richness rises about 2.6%. Keeping exactly three activity tags per qualifying game yields about 4.5%. Restricting to games with at least 100 reviews yields about 10.2%. Coverage conditions change the population, so these are robustness views rather than interchangeable estimates. Valve's official June 2020 announcement introduced a Steamworks Tag Wizard beta. That is consistent with the timing of a change in tagging practice, but the snapshot alone cannot attribute the full observed discontinuity to that tool. The main quantitative reading therefore emphasizes the more comparable recent cohorts instead of claiming a sudden creative expansion around 2020. ## Growth inside familiar labels Using each game's highest-ranked tag in the fixed activity set, the largest absolute additions from 2021 to 2025 are Adventure (+897), Action (+835), Strategy (+709), Horror (+521), RPG (+388), and Action Roguelike (+321). These six account for about 47.6% of net growth. Most are already broad, familiar categories. Shares nevertheless move: Horror rises from 2.5% to 4.6%, Action Roguelike from 1.1% to 2.6%, Idler from 0.2% to 1.3%, and Incremental from 0.5% to 1.3%. Generic Action falls from 16.4% to 13.6%, and Puzzle from 7.0% to 4.3%. These are primary-label shares, not all games possessing each feature. A more specific label gaining rank can displace a broader one without the game itself changing category in a deeper sense. 2,490 pairs in the 2025 cohort are absent from the 2021 cohort under the same first-five activity-tag definition. However, only 268 appear in five or more 2025 games and just eleven in twenty or more. Together they account for 4.5% of 2025 pair occurrences. The repeated additions include Desktop Companion + Idler, Management + Shop Keeper, Job Simulator + Management and Idler + Loot. Pairs involving Bullet Heaven illustrate the limit: the label can be absent from an older cohort even where a relevant activity existed. These are additional observed label combinations relative to 2021, not identified inventions or historical family trees. ## Do neighboring pitches become more alike? Equal-size 3,000-game samples compare each game's closest other game. For tags, the metric is Jaccard overlap; for short descriptions it is lexical TF-IDF cosine similarity in an explicitly English-oriented subset. Outcomes/review counts do not enter the similarity calculation. With different developer-credit strings, median nearest tag similarity is 0.421 in 2021 and 0.429 in 2025. Median nearest short-pitch similarity is 0.156 and 0.153. Removing shared publishers gives 0.417/0.429 for tags and 0.155/0.153 for pitches. Smaller random replicates and one-per-credit samples preserve the broadly flat recent pattern. All configurations compare equal numbers of games. The example audit includes exactly repeated puzzle descriptions and regional editions of Wolfenstein: Youngblood. Such cases demonstrate that repetition exists, but many high-similarity neighbors share a publisher or are variants. They do not justify calling a large share of the market clones. Conversely, different wording does not establish different gameplay. The supported reading is that many more games occupy familiar advertised territory, while the breadth and distribution of that territory change more slowly. That does not make those additional games valueless; execution, presentation, specific rules and audiences are not resolved by these measures. ## Reproduction and checks The fixed 181-label activity set, all-tag alternatives, coverage tables, exact rarefaction results, neighbor samples, example pairs and scripts are retained. The figure was generated locally; no store images were downloaded. Mechanical verification passed 48 checks for identities, count/entropy recounts, rarefaction bounds, equal comparison-pool sizes and score bounds. The original snapshot and paused bulk-network policy are unchanged.
