Atlas of Doors is a roguelike deckbuilder that runs in a browser tab, with no download and no account. It is also an experiment in building a game the way you would build a product: every door opened, every card played and every run abandoned is a logged decision, so nothing about how this game is doing has to be a matter of opinion. A month later there are a million of those decisions on record. Here is what they say, including the parts I would rather they did not.
Most small games ship and then guess. This one was built decision-tree-native on purpose: the whole game is cards, doors and nodes, which means every single interaction is already a structured event. We did not add analytics later. The telemetry went in at commit one, because the data is not a reporting layer on top of the game, it is the thing that tells us which version of the game to keep.
Not pageviews. The log is the shape of a decision: what options a player was shown, and which one they took. Multiply that by 11,165 started runs and you get something genuinely rare for a game this size, which is a complete, labelled record of how thousands of strangers make choices under pressure.
| Signal | Count | What it tells us |
|---|---|---|
| Events logged | 1,058,047 | Every decision point, both options shown and option chosen |
| Unique devices | 10,171 | Anonymous device ids, not accounts. No sign-up wall anywhere |
| Sessions | 16,496 | More sessions than players, so a share of them come back |
| Runs started | 11,165 | A run is a full descent, not a tap |
| Runs with a full decision trace | 8,020 | The north-star. Fully labelled play, start to finish |
This is the part I am quietly proud of. A month ago this game had no players, no audience, no wishlist and no press. Getting to ten thousand devices is not a mystery any more, we know how to do it and we know roughly what it costs. That is a real piece of know-how, and it is separate from the harder question, which is whether they stay. More on that shortly, because the answer is "not enough of them yet".
Totals go up simply because time passes and we keep buying traffic. Ratios do not. So the four charts at the bottom of that dashboard are the ones I actually read, because they divide by the player count and therefore cannot be flattered by spending more money.
| Per active player | Now | Range over the window | Direction |
|---|---|---|---|
| Sessions / player | 1.62 | 1.12 to 1.76 | Climbing |
| Runs / player | 1.09 | 0.09 to 1.43 | Climbing |
| Completed runs / player | 0.82 | 0.00 to 1.24 | Climbing |
| Committed players | 2.7% (277) | 0.0% to 15.5% | Climbing |
These moves are not dramatic. Runs per player went from roughly 0.90 to 1.10, finished runs per player from roughly 0.70 to 0.90. Call it ten percent, steadily, on a base that keeps getting diluted by a firehose of brand new traffic. That last part is what makes it meaningful: we are pouring in hundreds of first-timers a day, which drags every per-player average down, and the averages went up anyway.
Every metric on our dashboard carries its own written definition, so it cannot quietly drift between versions and flatter us by accident. Here they are, goals included, with the ones we are failing marked as failing.
Of engaged human opens, meaning crawlers, link-unfurl bots and browser prerenders filtered out, 35% go on to start a run. It was 25%. That is real progress and it is nowhere near enough, because the other 65% are people who came all the way to the door and turned around. This is where the next block of work goes, and it is the highest-leverage square on the whole board for a reason I will show you in the funnel section.
4% against a 15% goal. Exact-day cohort retention: of the devices old enough to count, the share that came back the very next day. Mobile-web games generally floor around 10 to 15%, so this is not good, and lately it has been drifting down rather than up.
I have a strong suspicion about why, and it is not the build. Our current ad round is re-hitting an audience pool we have already worked hard. Those people open the game, recognise it, and leave, which is exactly what a re-served impression looks like in a retention chart. Pool depletion reads as a product regression if you do not split the data by campaign, so that is the next thing I split. Until then I am treating the dip as unproven rather than as a verdict.
From 1% to 2% across the window. The goal is 10%. Day-7 is the lever worth the most over time, because a change that lifts day-7 lifts it for every future cohort as well, not just for the people currently in the game. Doubling in three weeks is a genuinely good trend. I will be honest about the caveat in the predictions section rather than pretending it extrapolates cleanly.
535 first-seen devices a day, for about $48 of daily ad spend. Note there is no goal set on this one, deliberately. We are not optimising for acquisition volume and we do not want to. We already ran a campaign that delivered more players more cheaply, and those players dropped off faster and enjoyed the game less than a smaller, better-matched group did. Buying a bigger number is easy. Buying a better number is the whole job.
4.5 minutes against a 10 minute goal, measured only over sessions that actually started a run. The median web game runs 5 to 6 minutes, so we are just under par and climbing.
9% of all devices show up on two or more separate days, against a 15% goal, and it has been rising the whole window. This is a looser measure than day-1: it does not care whether you came back tomorrow or next Tuesday, only that you came back at all. For a game with no push notifications, no email list and no install prompt, every one of those returns is somebody choosing to type the address again.
Runs per player sits at 1.1 and climbing, which is the "one more run" loop doing its job. Repetition is the whole hypothesis of a roguelike. If people replay, they stay.
Everything in the last section pools two completely different populations: people who played the game, and people who looked at a home screen and left. Since roughly two thirds of arrivals never start a run, the pooled numbers are mostly measuring the bounce. So here is the same data restricted to people who actually pressed play.
| Metric | Everyone who opened | Started a run | Goal |
|---|---|---|---|
| Day-1 retention | 4% | 9% | 15% |
| Day-7 retention | 1% | 3% | 10% |
| Returned within 2+ days | 9% | 18% | 15% Met |
Day-1 more than doubles. Day-7 triples. And return-within-two-days lands at 18% against a 15% goal, which is the first target on this dashboard we have actually met.
The honest caveat: this is a selection effect, not a controlled experiment. People who choose to start a run were always going to be more interested than people who did not, so converting a reluctant visitor will never produce a player as good as a naturally motivated one. But the gap is far too wide to be only selection. There is a lot of retention sitting behind that door.
This is the whole per-player journey, nested, built for one purpose: to find where people go. Red is a big leak.
Six out of ten people who see the home screen never see the next screen. This is not a game problem, nobody has played anything yet. It is a first-paint, comprehension and intent problem: what is this, why should I care, and what do I press. We have made real progress here already, the home screen has been through several rounds of work, but a month-long window averages the old versions in with the new ones, so the chart understates where we actually are today. The last seven days look better than this picture does.
The second big one, and much more interesting, because these people did play. They picked an element, opened a door, and then a bit less than four in ten stopped before the second floor. Something in that first fight or first reward is losing people who had already decided to try. We have improved it and it is still the second-largest hole in the game.
| Of the 3,665 who started a run, how many ever... | Count | Share |
|---|---|---|
| Completed a run, any ending | 1,586 | 43% |
| Reached the boss fight | 1,220 | 33% |
| Played a second run | 1,109 | 30% |
| Won a run | 1,057 | 29% |
| Came back another day | 918 | 25% |
| Played a third run | 622 | 17% |
| Shared a result | 44 | 1% |
Thirty percent play a second run and a quarter come back on a different day. That is a game people like. One percent share anything, which is a game nobody is telling their friends about, and that is a separate problem for a separate month.
A dashboard will never tell you that somebody enjoyed themselves. For that you have to read your messages.
"I came across the ad for Atlas of Doors. It looked interesting so I clicked. And I have been playing it for like 3 days now. It has been a great time kill. Just wanted to let you know I have been enjoying it immensely and nearly completed the codex." A player, unprompted, three days after an ad click
The second one is better, because it is not a compliment. Somebody tracked me down specifically to file bug reports, at midnight, and then stayed to argue about card balance.
"I really love the game, and I have some feedback for the developer as to errors/glitches as well as what I would love to see implemented. Because the game leaves me wanting more." The most useful message we have received
People who complain in detail are people who care. That message produced a patch: no-path runs no longer snap you back onto your previous path, Pacts persist properly, and the Descent arrows are actually pressable. The card balance point turned into a full rebuild, which became Season II.
We opened a forum so there is somewhere to talk that we do not control the tone of. Announcements, a changelog thread per version, dev write-ups like this one, and player-started discussion. A private Discord is on the plan for later. Being straight about it: we are moving slower on community than we probably should be, and it is a deliberate trade rather than an oversight. Building in public costs time that currently goes into the build.
Content velocity is the one advantage a very small team actually has, so we spent it.
Every card and every effect rebuilt from scratch. Not tuned, rebuilt: new keywords, new interactions between paths, and a deliberate design spine where each archetype owns a complete internal loop instead of borrowing bits from the others. This is the answer to that "some cards are incredibly powerful" message.
12 new monsters and 3 new biomes. Before this, the eight floors of the descent were one corridor wearing different monsters. Now they are three distinct places, each with its own foes, its own room and its own sound.
Both landed inside the measurement window, which is worth remembering when reading any trend line above: the game people played on Jun 27 is not the game people played on Jul 18.
Roughly $2,000 out. Five dollars in. Here is exactly where it went.
About $1,250, of which roughly 70% was image generation, 10% video, and 20% ordinary coding assistance. Most of that image spend was simply wasted: we generated at resolutions far higher than the game will ever display, with no batching and no optimisation pass, which is a beginner's mistake that costs real money at volume.
We now burn several times less for the same output. Icons are generated five to an image in one call instead of five calls. Every generated asset goes through an optimisation pipeline that produces a mobile-sized WebP. We measure the target aspect ratio before generating rather than cropping afterwards. The tools got better because the bill got embarrassing.
For the curious: Gemini (Nano Banana Pro) for images, Veo 3.1 for video, and Claude Code on a Max subscription for the engineering, run as agentic loops rather than as a chat window.
$845.84 total across five rounds. Our hard monthly ceiling is $1,250 and we are comfortably under it. That budget is not trying to build a business, it is buying answers: enough real players to read retention honestly, enough traffic to A/B test against, and enough reach that the game has an audience to talk to.
The game is free, has no payments wired in, and does not ask. Somebody tipped five dollars anyway. Measured against a $2,000 burn this is a rounding error, and I am not going to pretend otherwise. But someone choosing to hand money to a free browser game in its first month, with nothing to buy and nothing to unlock, is the cleanest willingness-to-pay signal available at this stage. It is the number I was happiest to see.
There is no monetisation in the game today and that is deliberate. Free removes every excuse not to play, which is exactly what you want while the thing you are optimising is retention. But we are testing the distribution paths now, because most of them take longer to open than to use.
Our first version went into CrazyGames' Basic Launch and started well. Two weeks in, it is not going to clear the bar. Playtime scores 1 out of 5 (5m34s desktop, needing 10m1s for even the next rung), gameplay conversion scores 1 (47.56%, needing 61.24%), and retention scores 3 (4.17%, needing 5.29% for a 4).
A newer version is already uploaded and waiting on review, and several further improvements are built but unpublished. In any failure there is a lesson, provided you actually read it.
Web portal distribution, revenue share. First submission grading below bar, newer build submitted.
The publishing process is long, so this one gets started sooner rather than later purely for calendar reasons.
Same logic, higher bar, later start.
Including a Steam Deck build. Wishlists compound, so opening this early has value even if the release is far off.
Both are embedded-play platforms where our audience already is. Availability and regional restrictions still being checked.
New archetypes, events and seasons, sold as packs. Every card and event in this game is already JSON with a pack id, so the engine has been built for this from day one.
Card backs, door skins. Cosmetic only, always with a visible real-money price, never affecting odds or difficulty.
Fits the season cadence we are already shipping at.
The goal is not revenue this quarter. The goal is to get an early, honest read on whether anyone would pay, because how you would sell a thing changes what you build. We are not rushing it, and we are not going to wire payments into a game that has not earned its retention yet.
Five rounds of Reddit ads in a month. Round 5 is live and running the winners from Round 4 for the next seven days, after which we re-cut based on whatever the ad centre tells us.
The operating principle is uncomfortable but it has held up every round: the ads that work today will not work tomorrow. Audience pools deplete. Creative fatigues. A hook that beat everything in round three can lose to a worse hook in round five simply because the people it appeals to have already seen it four times. So we do not get attached: not to an ad, not to a creative angle, not to a channel, not to a tool.
One hard-won reading lesson from round four: better hooks bought us better first days and precisely zero extra second weeks. Round 3's winners lifted clicks and did nothing measurable to retention at day 7, which came in at 1.14%, statistically identical to the campaign those hooks had beaten. Acquisition cannot buy retention. You can buy a bigger top of funnel all day long, and the game still has to do the rest.
We are staying on Reddit for now. Not out of loyalty, but because the cost per player is satisfying and the audience fit is good, so there is no reason to spend attention on a second channel while the first still has room. Other channels come later, once there is a retention number worth pouring traffic into.
Writing these down so they can be checked against reality next month rather than quietly forgotten.
Pooled day-1 is roughly activation multiplied by player day-1. At 36% activation and 9% player day-1 that is about 3.2%, which brackets the 4% we see. Push activation to the 50% goal, change nothing else about the game, and pooled day-1 should land near 4.5 to 5%.
Even at 100% activation, 9% player day-1 caps pooled day-1 at 9%. The 15% goal is unreachable by fixing the front door alone. It requires the first fight and first reward to get materially better too, which is the second funnel leak.
Day-7 went 1% to 2% in three weeks. The tempting extrapolation says 10% in about seven more weeks. I do not believe it. Early doubling off a tiny base is the easiest doubling there is, and it decelerates. My actual call: 3 to 4% pooled by the end of next month, which would be excellent and is not 10%.
Session length responds to build quality within days. Retention responds within weeks, and only for cohorts acquired after the change. If both are moving, the fast indicator moves first. If session length climbs and retention does not follow within two weeks, my model is wrong and I will say so.
157 to 277 in three weeks, while daily new-player volume stayed roughly flat. If activation improves at all, the committed count should reach 450 to 600 next month. This is the cohort I would trade every vanity metric for.
1% share rate will not improve by itself, because nothing in the game currently gives a player a reason or a moment to share. This will still be 1% next month unless it becomes somebody's actual project, and it is not scheduled to be.
Six in ten people leave at the home screen. This is the largest single block of value on the board and everything downstream is gated by it. Expect first-paint work, clearer comprehension of what this game is within five seconds, and a shorter path from arriving to playing.
The 38% who quit between the first door and floor 2 had already decided to play. Whatever is losing them is a first-run experience problem, not a difficulty problem, and it is the second-biggest hole in the game.
More content, and specifically new bosses paired to the descent rungs, currently being designed. The top of the ladder should fight back harder than it does.
Before blaming the build for the day-1 dip, prove it is not an acquisition mix shift. This is a day of analysis that could save a month of chasing the wrong problem.
Google Play first, because the publishing pipeline is long and there is no reason for the calendar to be the bottleneck later.
Which is not a filler item. A month in, with a million events logged, two seasons shipped and five dollars earned, the thing that determines whether there is a second month is whether the people building it still want to.
Because the numbers are the game. Atlas of Doors logs every decision so that design can be argued with evidence instead of taste, and it would be dishonest to use that data to make the game better and then only show you the flattering half. Day-1 is 4% against a 15% goal. Six in ten people leave before they play. One percent share. If you have played this game you are somewhere in these charts, and if you closed the tab, the reason you closed it is the single most valuable thing anyone could tell us right now.