Module 5 · The HD era and indie (2005–2012)

Steam, wishlists and launching

Making a game is no longer the hard part — getting it found is. Steam is not a shelf but a visibility algorithm: 19 thousand releases a year, and about half collect fewer than 10 reviews. Visibility is won before launch — with wishlists, a demo, and a first week that decides everything.

~17 min💰 business + 🎲 algorithm

The gist in 30 seconds

Steam gives you a global shelf, the tools (Steamworks, free) and an audience for a cut: 30% of the first $10M, 25% from $10–50M, 20% above that. But the scarce resource is not shelf space, it is attention: 19,112 games shipped in 2025 (~52 a day), nearly half with fewer than 10 reviews, and >40% did not even make $1000. Visibility is run by the algorithm: every new release gets its "15 minutes of fame" (~25,000 impressions in the first day or two), the algorithm measures conversion — and then either amplifies (impressions→sales→reviews→more impressions) or buries. The fuel for that flywheel is wishlists: they bank targeted demand, fire a notification cannon on release day, and drive the day-1 spike that the algorithm reads as "this converts → amplify". The "Popular Upcoming" threshold is around 7000 wishlists. So marketing is not something after release but months before: bank wishlists (a demo, Next Fest, a community), and the launch is a harvest in a window that will not come again.

The mechanism: visibility as an algorithm, not a shelf

When Steam launched (2003, initially as DRM for Half-Life 2), being listed = being visible: there were hundreds of games and the store was curated by hand. Today that is inverted. Greenlight (2012) and then open publishing removed the gate — and the storefront drowned: 19,112 releases in 2025 against ~9,700 in 2020. At 50+ games a day no human being browses the catalog — a recommendation algorithm does it for you. The shelf is infinite and free; impressions are scarce.

The economics of the shelf — what the 30% buys

Valve takes a marginal cut: 30% of a game's first $10M of revenue, 25% of the $10–50M slice, 20% above $50M (the tiers date from 2018). A game grosses $12M — Valve takes:

0.30 \cdot 10 + 0.25 \cdot 2 = 3.5 M $ (effectively 29 %)

For that: hosting and bandwidth, Steamworks (achievements, cloud saves, matchmaking, Workshop) for free, refunds and support, anti-fraud, and — most importantly — an audience and a discoverability algorithm. Against the 40–50% of 2000s retail, 30% looked like salvation; against the Epic Games Store's 12% today, it looks like grounds for an argument about platform rent (Steam's PC share is ~75%).

The visibility machine — positive feedback

The heart of Steam is an algorithm that bets on early success. The rough mechanics of a launch:

A new release is granted a starting budget of impressions — on the order of 25,000, its "15 minutes of fame" in the first day or two (this is forced exploration: even a total unknown gets a shot).
The algorithm measures conversion (sales / clicks) and reviews.
Converts well → more impressions → more sales → more reviews → more impressions still. Converts badly → visibility is cut and the game sinks.

This is rich-get-richer: the "visibility → sales → reviews → visibility" flywheel either spins up or stalls in the first 48 hours. Hence "the first week decides everything": the window is narrow, and the algorithm does not grant a second launch (a long tail like Among Us is rare luck with streamers, not a plan).

Wishlists — the fuel of the flywheel

A wishlist looks like a bookmark, but it is a warm start for the algorithm and a piece of targeted demand. It does three jobs at once: (1) on release day everyone who added it gets a notification (a free marketing cannon); (2) that volley produces a day-1 sales spike that the algorithm reads as "this converts → amplify"; (3) banked wishlists open the "Popular Upcoming" shelf (a threshold of ~7000). So all the marketing shifts before release: a demo, taking part in Steam Next Fest (a demo showcase: people who wishlisted get a push when the demo goes live), a community — all of it to bank wishlists by day one.

A worked example. You have banked $W$ = 10,000 wishlists; in the first week $c$ ≈ 10% convert:

S_{1} \approx c \cdot W = 0.1 \cdot 10000 = 1000 sales on launch day

That spike is not merely $20k of revenue but a signal: the algorithm sees high day-1 conversion and pours in more impressions. The same 1000 sales spread over a year would never start the flywheel. That is why it is "bank wishlists → fire them all on release day" and not "upload it and wait".

🕹 What to open, and what to notice

Here what you "play" is not a game but the storefront — the mechanics are visible right in the Steam interface and on SteamDB. From a single success to the system.

Balatro 2024 · a Next Fest demo → wishlists → a hit

A solo developer: the demo of a roguelike poker game took off at Steam Next Fest (Oct 2023), banked wishlists over the following months — and the release (Feb 2024) landed straight into the flywheel, becoming one of the year's biggest hits and a GOTY nominee. A textbook case of "bank wishlists through a demo before launch".

🎮 Watch: open the Balatro page and its demo — notice the separate store pages and the demo's own reviews. On SteamDB look at the followers/wishlist proxy curve over time: growth runs up to release, and on launch day sales jump vertically. That is banked demand firing.

Vampire Survivors 2022 · early access at ~$3 → word of mouth

It shipped cheap in Early Access and spread through streamers and word of mouth — a low price plus insane conversion spun the algorithm up from below. A counter-example to the expensive PR launch: sometimes the flywheel starts on price and virality rather than budget.

🎮 Watch: compare its price tag and review count (tens of thousands) with the median 2025 release (<10 reviews). Notice how a cheap impulse purchase pushes up conversion — fuel for the same algorithm, but through price rather than wishlists.

The storefront as a whole Next Fest · "Popular Upcoming" · <10 reviews

The system becomes visible once you look at it as an algorithm. 19,112 releases in 2025, about half with <10 reviews, 2229 with no reviews at all; the flywheel found a handful and buried the rest at launch.

🎮 Watch: go to Steam Next Fest and scroll through demos — notice the churn in "New & Trending". Open "Popular Upcoming" (you need ~7000 wishlists to get there). Then find a recent release with 3 reviews and ask: where were its wishlists, its demo, its community? Almost always: nowhere. Visibility was lost before the release button.

Deep end · economics: the unit economics of a launch and why >40% never make $1000skippable

The numbers of an indie launch are merciless. The cost of entry is $100 (the Steam Direct fee), refunded once the game grosses $1000. In 2025 >40% of releases did not even reach that $1000 — meaning they formally failed to recoup the fee itself, never mind years of development.

Where the revenue leaks

Out of a nominal $1000: −30% to Valve = $700; −refunds (the 2 hours played / 14 days window, ~5–15%); −regional pricing (key markets are several times cheaper); −taxes and payment fees. "They paid $20" ≠ "I received $20". So the real survival threshold is not $1000 but tens of thousands of copies, and nearly all revenue arrives in the first week or two (often 50%+ of lifetime), while the flywheel is spinning and before the discount wave hits.

The shape of the distribution

The market follows an extreme power law: a handful of hits take almost all the revenue and the median is near zero. 19k releases a year is not "a lot of competitors" but noise in which the signal drowns; add AI slop (~8000 games flagged with AI content in H1 2025 against ~1000 in all of 2024, ×8) and the signal-to-noise ratio of the storefront drops further. The conclusion: distribution is not "uploading" but a separate product with its own budget and schedule.

Deep end · the algorithm: Steam visibility as a recommender systemskippable

Underneath, Steam's algorithm is an ordinary recommender with an explore/exploit dilemma and a cold-start problem.

Explore vs exploit

The "15 minutes of fame" (~25k starting impressions for everyone) is forced exploration: the system needs to collect signal about a release that has no history yet (like ε-greedy, or a bandit pulling an arm it has no statistics for). After that comes exploitation: impressions flow where conversion is higher. Each game = an arm of the bandit, impressions = pulls, a purchase = a reward.

Cold start and the feedback loop

With no reviews and no history there is nothing to rank on — that is cold start. Wishlists and demos are the side information / prior that treats it: they provide a demand signal before any reviews accumulate. And "amplify whatever already converts" creates popularity bias / a feedback loop (rich-get-richer): the algorithm makes the rich richer, like the YouTube/Spotify recommenders. That is not a storefront bug but a well-known RecSys pathology — and the reason "a good game" is not sufficient: without an early signal the system cannot see you well enough to learn that you are good.

Analogy

Steam is a casino with 19,000 slot machines and one spotlight, which the owner points at whichever machine just paid out. Your wishlists are the crowd you brought with you: the moment the doors open they all rush your machine, it "pays out" (day-1 sales), the spotlight swings onto you, passers-by drift over, more payouts, a brighter spotlight — the flywheel. Bring no crowd to the door → no early payout → the spotlight never finds you → you are one of thousands of dark machines. The quality of the game is the machine's payout probability; but the spotlight is won at the entrance.

Why it matters

Making a game is no longer the bottleneck; being found is. In a market of 19k releases a year where half collect <10 reviews, distribution is a positive-feedback algorithm that has to be charged before launch with wishlists and a demo. Understanding the visibility machine — cold start, the first-week flywheel, the wishlist as a warm start — is the difference between "shipped into the void" and "shipped into the spotlight". And these are exactly the laws of recommender systems that govern any attention market you will build or compete with.

🔁 Beyond games — where this transfers

The lesson is discoverability in a crowded marketplace: cold start, positive feedback, and the fight for attention as the scarce resource.

ML / AI (your domain): Steam's algorithm is a recommender system in its purest form, and all its pathologies are yours. Explore/exploit and multi-armed bandits: the "15 minutes of fame" = forced exploration of a new arm; ranking by conversion = exploitation. Cold start: no history, nothing to rank on; wishlists/demos = the side features and priors that treat it (a content-based bootstrap over collaborative filtering). Feedback loops and popularity bias: "amplify the popular" makes the rich richer and collapses diversity — the known fairness/feedback-loop problem in RecSys (the same one that afflicts the YouTube/TikTok feeds). And the ×8 in AI slop on the storefront = adversarial content in a recommender: a spam arms race you solve with filters and trust signals. If you build ranking, or an agent competing for distribution, this is your subject matter, not a metaphor.

Platforms / marketplaces: the App Store, Amazon, YouTube, Spotify — the same laws: the launch window = the exploration budget, reviews/ratings = the signal, early traffic decides; "SEO/ASO" = engineering against somebody else's ranking algorithm.

Product / GTM: "build the audience before launch, the release is the harvest" = waitlists/betas/devlogs; the first week as the point of maximum leverage; distribution as a separate product with a budget, not "we'll push it afterwards".

The principle: in any channel where attention is distributed algorithmically, the winner is not the best but the best with an early signal. Charge the flywheel before launch; the window is narrow and it opens once.

🔧 Run it and poke at it — on your home machine

What to look at is above (🕹). This part is about taking the visibility machine apart with instruments.

🔧 Poke at it (debug) ~40 min, SteamDB + Steam

Take 3 games (a hit, a middling one, a stillborn one) and compare their curves on SteamDB: follower dynamics, the date and height of the day-1 spike, how fast reviews accumulate. Then take apart their store pages as CRO: the capsule, the first frame of the trailer, the tags, "Popular Upcoming", demo or no demo. Form a hypothesis: what was wrong with the dead one before release?

🧪 Test it (with marketer eyes) ~15 min

Go to the current or last Next Fest, pick 5 demos and predict which will take off at release, based on: the quality of the capsule/trailer, the demo's review count, how clearly the game's "verb" reads in 5 seconds. Then sketch a mini launch plan for an imaginary game: how many months to bank wishlists, where the first thousand come from, how not to scatter the day-1 volley.

Checklist: compared the visibility curves of 3 games on SteamDB; took a store page apart as CRO; explained a "death before release" as a missing early signal; sketched a wishlist launch plan.

Connections

foundation

Engines — that is what you built the game with; now you have to push it through the storefront. Making ≠ selling.

contrast

F2P economics and gacha — there the money is squeezed out of people who are already inside (monetization); here it is about getting on the radar at all (distribution). Two different funnels.

adjacent

Virtual economies — there it is the player↔player economy inside the world; here it is the platform↔developer economy around it (rent, the cut, visibility).

Questions worth asking

19k games a year, half with <10 reviews — how do you break through at all?

Not at launch but before it. The algorithm only amplifies what has already shown day-1 traction — which means you have to bring the traction with you: months of wishlists, a demo, Next Fest, a community, press. The launch is a harvest of banked demand, not its beginning. "A good game will find its audience" is false in 2025: a good game with no early signal is invisible — the system never gets the chance to learn it is good. That is why distribution has become a separate product with its own budget and schedule, often larger than the one for gameplay.

Why do wishlists matter so much — they're just bookmarks?

Because they are charged, targeted demand doing three jobs at once: on release day everyone who added it gets a notification (a marketing cannon you do not pay for); that synchronized volley produces a day-1 spike the algorithm reads as "high conversion → amplify"; and accumulating them opens "Popular Upcoming" (~7000). In effect a wishlist is a warm start for the recommender: you deposit a demand signal in advance so that cold start does not bury you before your first review. The same additions spread across a year would not work — you need the synchronized shot.

Steam takes 30% with ~75% of the market — is that rent or a fair price?

There are arguments on both sides, and the debate is real. Against Valve: 30% is a lot (Epic takes 12%), a near-monopoly is what lets them hold the rate, and the cut is taken even on sales to your own audience that you brought in. For Valve: the price includes hosting/bandwidth, free Steamworks, refunds, anti-fraud, an enormous built-in audience and — crucially — discoverability; the tiers (25/20%) reward hits; and leaving for your own storefront (as EA did with Origin) has historically lost more sales than it saved in commission. The core of the argument is what Steam's distribution and audience are actually worth: if it is close to nothing, 30% is rent; if it is half your sales, it is payment for a result. The answer depends on whether you could have sold those same copies without Steam.

Why the first week specifically, rather than "quality will pull it through eventually"?

Because of the algorithm's positive feedback: day-1/2 conversion determines whether you get amplified or cut, and that compounds (visibility→sales→reviews→visibility). Miss the window and the flywheel never starts, and there is almost nothing left to lift a dead listing later: the "15 minutes of fame" are not granted twice. Long tails happen (Among Us took off two years later thanks to streamers) but that is rare luck, not a strategy — it is usually triggered by an external burst of attention you do not control. Plan for the narrow window, not for a miracle.

AI slop grew ×8 in a year — will that kill Steam's discoverability?

It damages the signal-to-noise ratio of the storefront: the more cheap content there is, the harder it is for both algorithm and player to tell what is worth anything, and the more weight shifts onto pre-release trust signals — wishlists, demos, press, community, known creators. It is a classic adversarial arms race in a recommender (spam vs filters): Valve adds AI-content labels and trust signals, spam looks for ways around them. The practical takeaway for a developer: curation and reputation before release get more valuable, and "just upload it and hope for the algorithm" works worse and worse.