Racing Someone Who Isn't There: What an Empty Matchmaking Queue Taught Me
Pick two articles and race between them. No account needed.Play WikiRacing
When I added ranked duels to WikiRacing in the summer, I pictured a queue: you press the button, someone somewhere presses it too, and thirty seconds later you are both racing from the same article to the same target. That is how it works on paper. In practice, the queue was empty most of the time.
It wasn't completely dead, which is the interesting part. Since August the database has recorded 155 ranked matches, 121 of which ended with a rating change. But look at who played them and the picture changes.
155 duels, 39 players, six friendships
Those 155 matches were played by 39 different accounts, split into 21 distinct pairs of opponents. Six of those pairs account for 119 matches, a little over three quarters. Two Polish players met 39 times in a single week at the end of October. Two French players did 29 duels in eight days. Seven pairs only ever met once, and those seven matches are the closest thing I have to the original idea: two strangers who happened to press the button at the same moment.
So ranked works, but mostly as a scheduling tool for friends. You message someone, you both open the site, you both queue, and the queue matches you with each other because nobody else is in it. Fine. It just isn't what a stranger arriving on the site for the first time experiences. That person sees a spinner, then a message saying no opponent was found.
The traffic is there. The timing isn't.
My first instinct was "not enough players." The numbers say otherwise. Over the last 30 days, 69% of all 15-minute windows had at least two races starting in Spanish. In German it was 62%. People are on the site at the same time, constantly.
And yet the launch log contains exactly two ranked starts in Spanish, ever. Being online at the same time is not enough: two people have to want a rated duel, in the same language, within the same couple of minutes. Each of those conditions cuts the pool. Multiply them together on a site where most visitors play a daily challenge and leave, and you get a queue that is technically open and practically empty.
I tried the cheap fixes first. The queue timeout went from 2 minutes to 15. A live counter now shows how many people are waiting. Those helped friends who had agreed to meet. They did nothing for the lone visitor at 2 p.m. on a Tuesday.
Borrowing the oldest trick in racing games
Racing games solved this a long time ago. Mario Kart 64 already let you race a translucent copy of your own best lap in time trials. Later games let you download other players' ghosts. Nobody needs to be online for that to work: the opponent is a recording.
WikiRacing has been sitting on thousands of recordings without using them. Every finished race is saved with its full path and its total time. So since November 7 there is a "Race a ghost" button under the ranked one, and another on the matchmaking screen so you can bail out of an empty queue without losing the moment.
Press it and the game picks a real finished race from the last couple of weeks in your language, puts you on the exact same start and target, and shows the other player's progress next to yours as the clock runs. When you reach the target, the result line tells you whether you beat them.
What a ghost actually is, and what it isn't
There's one thing I want to be upfront about, because it would be easy to oversell. The game stores the path of a race and its total duration. It does not store when each individual click happened. So the ghost's progress during the race is an approximation.
A real example: on November 5, a French player finished that day's daily challenge, Napoléon Ier to Bitcoin, in 3 clicks and 59.7 seconds, going through "Liste des monnaies de la France" and "Système monétaire". Race against that run as a ghost and you will see it advance roughly every 20 seconds, perfectly evenly. The real player probably spent 40 of those seconds reading the Napoleon article and then clicked twice in quick succession. Nobody knows. What is exact is the finish: 59.7 seconds. That's the number you are actually racing, and it's the only one used to decide who won.
The pick isn't random either. A run only becomes a ghost if it took between 2 and 15 clicks, lasted at least 15 seconds, averaged at least 3 seconds per click and finished in under 15 minutes. That filters out one-click flukes, abandoned tabs, and the occasional run that is simply too fast to have been played by a person reading pages. Your own runs are excluded, and when there are enough candidates, runs by players with a username are preferred, so the result line can tell you who you beat instead of "an anonymous player." Before the race starts, the game also checks that both articles still exist under the same title, because Wikipedia moves pages and a ghost aimed at a renamed target would be unbeatable.
Why ghosts don't touch your rating
I went back and forth on this one, and landed on no. Three reasons.
First, the person behind the recording never agreed to a rated match. They played a casual race a few days ago. Taking points off them, or handing you points for beating them, would turn their old game into something it wasn't.
Second, ranked duels go through extra checks on the server before any points move. A ghost comes from the general games table, which accepts anonymous saves. It is good enough to race against. It is not good enough to move a number that people care about on a leaderboard.
Third, a ghost doesn't react. In a live duel you see your opponent pull ahead and you take risks. The ghost will finish in 59.7 seconds whatever you do. That's a different kind of race, closer to a time trial than a duel, and mixing the two in one rating would make the rating mean less.
Ghost races are saved with their own mode tag, so they show up in your history and in the recent activity feed as what they are.
Your own record is the toughest ghost
The same machinery now handles something players were already doing. If you replay a pair you have already won, your best time on it shows up next to you as "Your record." Signed-in players get their fastest saved run. Anonymous players, who play about 78% of all games, get the record stored in their browser.
Replaying is common. Among signed-in players, about 18% of solo games and 27% of daily runs are on a pair the player has already won. So I looked at what happens on those replays, and the curve is nicer than I expected.
On a second attempt, players beat their first time 81% of the time, and when they do, they are typically about 2.6 times faster. No surprise: you know the route now. On a third attempt, 68% still improve. On a fourth, 64%. From the fifth attempt onward, only 26% manage to set a new best.
That is roughly the shape of any skill you practise. The early gains are the route itself; after a few tries you are fighting over seconds of reading speed and hesitation. Which makes the "Your record" ghost an easy opponent the first time you replay a pair, and a genuinely hard one by the fifth.
How many ghosts are there?
It depends heavily on your language. Over the last 14 days, the pool of races that pass the filters above was 2,837 in German, 2,600 in Spanish, 1,432 in French, 1,070 in English and 783 in Turkish. Polish has 413, Portuguese 319, Russian 293. At the bottom, Swedish, Chinese and Vietnamese have a dozen or so each, and Czech has 7. In those languages you may see "no ghost available" for a while, which is honest, if not exciting.
The feature went live yesterday, so I can't tell you yet how people use it, whether they win more than they lose, or whether they go back to the ranked queue afterwards. I'll look again in a few weeks.
What I do know is that the duel I imagined in the summer needed two people to show up at the same minute. A ghost only needs one person to have shown up at some point in the last two weeks, and on most days thousands already have.


