Methodology

Sales estimate

Model

PGS Sales Model, current generation.

Data inputs

Steam review totals and positive-to-negative ratio, genre bands, tag dampeners for known outlier categories, launch cadence and days-since-release, regional review-language mix (used to detect regional-market titles that under-report reviews), Early Access graduation signal, viral-peak anchoring for mega-launches.

Validation

Continuously cross-referenced against developer-reported sales figures. Every new dev-published number is a calibration anchor.

Confidence

High for released commercial titles more than 30 days post-launch.
Medium for pre-release estimates and titles inside their launch window.
Lower for free-to-play viral hits where community gifting and platform-cross-play distort review counts.

Sales estimates should be treated as modelled ranges rather than audited sales figures. They are designed primarily for comparative analysis across titles and for estimating the likely order of magnitude of an individual game's performance.

Wishlist estimate

Model

A segmented model combining Steam wishlist rank and follower behaviour.

Data inputs

Steam public wishlist rank when disclosed by Valve, follower count, first-tag genre multiplier, Early Access flag, follower conversion rate for rank-stable titles, segment-straddle detection for fast movers.

Validation

Cross-checked against developer-disclosed wishlist totals whenever those become public.

Confidence

High for titles inside Steam's disclosed rank window.
Medium for long-tail ranks and titles with a stable rank but rapid follower growth.
Advisory for unranked titles with fewer than a few hundred followers.

Wishlist estimates for event-window debuts (Gamescom, SGF, Next Fest, TGA) tend to run below the eventual total because these games arrive via discovery outside a normal follower base. We do not retune the model against these events since it would drag accuracy in normal weeks.

Potential Hits momentum

Model

Composite score, ranked output.

Data inputs

Follower velocity, press-coverage volume (major gaming outlets), live Twitch presence, Steam public wishlist rank, Google Trends search interest, franchise-recognition, Discord community.

Validation

Retrospective, not predictive. We track how the score at T-30 days correlates with subsequent launch-week peak concurrent users. The board self-corrects as new launches age in and out of the window.

Confidence

Advisory ranking rather than a numeric forecast. Useful for spotting which unreleased and just-released games are showing signs the other lists do not surface.

Sudden viral hits remain a structural blind spot for any pre-launch model that reads Steam-native signals. PEAK, Schedule I and similar titles can go from very limited pre-launch signal to chart-topping CCU in weeks. We accept this and do not tune the model around exceptional viral events. However, the combination of inputs can still provides early indicators of breakout behaviour beyond the normal parameters.

Review word cloud

Model

Baseline-normalized term sentiment.

Data inputs

Around 500 English reviews per game, tokenised nightly. Terms are coloured by each term's positive share minus that game's own baseline positive share, so relative aspects surface rather than flattening against a fixed 50/50 threshold.

Validation

Not applicable. The cloud is descriptive, not predictive. It summarises what reviewers wrote, not what they will write.

Confidence

Higher for games with at least 200 English reviews. Lower for smaller review pools where individual reviewers can dominate the per-term positive share.