Roguelike deckbuilders are one of the most crowded corners of Steam right now. Before Next Fest 2026 we did something unglamorous — pulled the numbers on 149 of them and went looking for what separates signal from superstition. The honest answer surprised us in a few places. A caveat up front: this is correlation on a single snapshot, not a law of nature, and we'll flag where the numbers get thin.

1. Wishlists are brutally top-heavy

The median game in the set has 360 wishlists. The mean is around 6,000 — and when the mean sits 17× above the median, you're not looking at a bell curve, you're looking at a power law. The top 10% of games hold roughly 85% of all the wishlists in the field.

Distribution of wishlists across the 149 games (lavender = median bucket, blue = mean bucket). The percentile ladder shows how fast the curve climbs at the very top.
If you're benchmarking your page against the breakout you saw on your timeline, you're benchmarking against the 1%. The realistic Next Fest deckbuilder is a three-figure wishlist game.

2. Most "this tag sells" advice is noise

We tested every tag carried by at least eight games — 66 of them — for wishlist lift, and put a real statistical bar in front of each: a bootstrap confidence interval on the median that has to clear the dataset baseline before we'll call it real. Exactly three cleared it: Auto Battler (~4.1×), Sci-fi (~3.9×) and Turn-Based (~3.4×). The flashy outliers — the ones riding on a handful of split games — evaporated once we accounted for how noisy a median of a few skewed numbers really is.

Tag lift vs. the dataset median (log scale). ● solid = statistically stable; ○ faded = tentative (CI straddles the baseline). Green is above baseline, rose below.
With ~150 games and this much variance, a single tag almost never moves the median on its own. Tag roulette isn't a strategy.

3. Localization correlates hard — and it's mostly a trap

This one looks like a cheat code. English-only games sit at 0.53× the baseline median; games supporting German, Brazilian Portuguese or Traditional Chinese run 3–4.5×. Don't read it as "translate to win." It's almost entirely reverse causality — broad localization is a proxy for budget and ambition. The teams that can fund eight languages can also fund the capsule art, the trailer and the scope that actually drive wishlists.

Median wishlists by language support. Tellingly, the cheap-and-common Asian-language support in this China-heavy indie field barely moves the needle; it's the expensive Western localizations that track with the big totals.
The languages are a symptom of a serious project, not the cause of its numbers.

4. The crowded lane is the worst lane

We clustered the field by tag profile, then plotted competition (how many games share a cluster) against performance (the cluster's median wishlists). The pattern is clean: the Atmospheric / Dark cluster is the smallest (13 games) and the highest-performing (~1,200 median), while the plain "Roguelike Deckbuilder / Card Game" cluster is the biggest (42 games) and below average (~320). The most contested lane is also the least rewarding.

Each dot is a sub-genre cluster. Up and to the left is the sweet spot: high interest, fewer competitors.
Tone and theme — the things that make a card game feel like a place — are where the differentiation, and the wishlists, actually live.

5. Check your data before you "discover" anything

The most useful finding wasn't about games at all. When we pulled follower counts and a third-party sales prediction, both showed a near-perfect 0.99 correlation with wishlists. Suspiciously perfect. It was arithmetic: for 128 of the 149 games, the follower count was exactly wishlists ÷ 12 — an imputed placeholder, not measured data. Only 21 games had genuinely independent followers, and on those the correlation fell to a normal 0.92.

Followers vs. wishlists (log–log). The grey diagonal is the artifact — 128 games sitting exactly on wishlists ÷ 12. Only the coloured points are measured.
If we'd trusted the headline figure, we'd have published a confident post about a relationship that was just our own input divided by twelve. Wishlists turned out to be the only independent success signal in the whole export.

How we did it

149 roguelike-deckbuilders from Steam Next Fest 2026. Because wishlists are power-law distributed, we used medians (never averages) and a bootstrap confidence interval to decide whether any "lift" was real or just small-sample noise. Tags and genres are treated as content metadata; wishlists as the one independent outcome. None of this is causal — read it as a map of where the field is dense, where it's thin, and which "best practices" survive contact with the numbers.