Where Players Give Up: What 81,000 Started Races Say About Wikipedia's Dead Ends

Since 11 July, players have started 81,082 races on WikiRacing. They have finished 42,069 of them. Roughly one race in two ends with somebody closing the tab, picking new articles, or hitting give up.

I had the first number for months and never looked at it closely, because a finished race is the satisfying thing to measure: a path, a time, a click count. An abandoned race leaves almost nothing behind. Since 22 September, though, every start also records its pair and language, which means the quits can finally be traced back to specific articles. So I did that.

One caveat before the numbers. A "start" is any launch, including a misclick or a restart after a bad first hop, so not every abandoned start is a frustrated player. That inflates all modes about equally, though, and the differences between pairs are far too large to be noise.

Not every mode loses the same people

• Campaign: 9,118 started, 6,718 finished (74%)
• Themed challenges: 14,301 started, 9,275 finished (65%)
• Daily challenge: 12,365 started, 6,135 finished (50%)
• Multiplayer races: 18,606 started, 8,588 finished (46%)
• Solo, manual or random: 26,490 started, 11,265 finished (43%)

The bottom of that list makes sense once you know how people use it. Solo is where the dice button lives, and the dice regularly hands out two articles that have nothing to say to each other. Rolling again is cheaper than slogging through a bad draw, so people roll. Multiplayer has its own exit: once someone else has crossed the line, a fair number of players stop racing for fourth place.

The campaign sits on top because its pairs were chosen by hand to be fair. Which is exactly why its second level surprised me.

Level 2 is a wall

Level 1 of the campaign is Bread to Wheat. Level 2 is Ancient Egypt to Mummy. On paper they're the same difficulty: one famous topic, one thing that obviously belongs to it. In German, Spanish, French and Polish (the four languages with enough campaign traffic to compare), 82% of level 1 starts are finished. Level 2 drops to 49%. In Spanish it's 45%, in Polish 41%.

And it isn't long. Players who do finish level 2 need a median of three or four clicks. Something on the very first page is sending people away.

That something is a missing link. The English article on Ancient Egypt links straight to Mummy. The French, Spanish, German and Polish ones don't. Spanish Wikipedia's Antiguo Egipto has 475 links and not one of them points to Momia. You arrive on a page about the civilization most associated with mummies in the world, scroll looking for the word, and it isn't there as a link. That's a strange feeling and plenty of people quit on it.

The finishers show the way around it. On French Wikipedia, 37 of them went through Pharaon first and 26 through Momification en Égypte antique. In German, 81 found Mumifizierung im Alten Ägypten, a page with only 130 links that does link to Mumie. In Spanish, Faraón was the favourite first click (25 games). The detour is short. It just isn't where your eyes go first.

I wrote that level. When I checked it, I checked that both articles existed in every language. I never checked whether the first one led anywhere near the second. Lesson learned, and I'll come back to it below.

One bridge, or none

The clearest contrast in the data comes from two daily challenges that came back through the rotation a week apart.

On 22 September French players got Pomme to Mécanique quantique (Apple to Quantum mechanics). It sounds brutal. It was one of the most finished pairs on the site: 218 starts, 167 wins, 77%. Of everyone who has ever won that pair, 142 out of 174 made the same first click: Isaac Newton. Everyone knows the apple story, the Pomme article mentions it (via Apple's first logo, of all things), and Newton's own article links straight to Mécanique quantique. One obvious bridge, and almost everyone takes it.

On 30 September German players got Wein to Quantenmechanik (Wine to Quantum mechanics). Same destination, similar-sounding start. 58 starts, 17 finished: 29%, the worst completion rate of any pair started at least 40 times. Winners scattered across twelve different first clicks: Ethanol, Temperatur, Österreich, Destillation, even Deutsche Nationalbibliothek. There was a two-click route the whole time, Wein to Temperatur to Quantenmechanik, and three players found it.

So the difficulty of a pair isn't really the distance between the two topics. It's whether a single, obvious bridge exists. When it does, people walk across. When it doesn't, they wander, and half of them go home.

The pages people back out of

Abandoned races aren't saved with a path, so for this part I only have the winners. Even among them, 10,194 of 42,069 finished races (24%) contain at least one step straight back to the article the player just left. I counted, for every article visited at least 60 times, how often the very next move was a retreat.

Top of the list on Spanish Wikipedia: Shōnen, with 16 retreats out of 61 visits. Almost all of them came from players in the manga challenge heading for Bleach, Hunter × Hunter or Shingeki no Kyojin. It looks like the perfect stepping stone. It's a short page of 4.5 KB with 31 links, and the only famous series it names are Dragon Ball, Naruto and One Piece.

Second, on French Wikipedia: France métropolitaine, 15 retreats out of 63 visits. This one is a trap with a very specific shape. It's one of the first links in the infobox of the France article, so people starting from France click it by reflex. Then they find a page of 72 links, against 2,846 on France itself. The destinations they were chasing (Vladimir Poutine, David Bowie, Chocolat) are nowhere near it.

Size isn't the whole story, though. New York City has 2,720 links and still sends 11% of its visitors straight back. Россия, on Russian Wikipedia, has 2,902 and 10%. A big page has a door to everywhere, but you still have to find the right one, and on a phone that means a lot of scrolling past doors that look plausible.

Spotting a dead end in three seconds

Most of what I'd tell someone comes straight out of those two examples.

Look at the scrollbar before you read anything. A tiny scroll thumb means a big page; a page that fits on two screens has very few exits. That isn't a reason to never click one, but it should be the exception you choose, not where you land by accident.

Be suspicious of a narrower version of the page you're already on. France métropolitaine, a "History of" page, a list article: they sound like a step forward and are often a step sideways into a smaller room. The broader article you came from usually has more links than the sub-page does.

Give a new page one screen. If nothing on it points toward your target's world (the country, the field, the era), go back. It costs one click. Staying costs you the next five, and the numbers above suggest it's also where a lot of people simply give up.

And if the start page doesn't contain the word you expected, don't search harder for it. Do what the Ancient Egypt winners did: take the closest person or institution (a pharaoh, a scientist, a ruler) and let that page do the work.

What the quitters taught me

Nobody writes a report saying "level 2 annoyed me". They close the tab. Which makes abandonment the most honest feedback a game like this gets, and also the quietest one.

From now on, when I check a new campaign level or daily pair, I'll look at more than whether both articles exist. I'll open the first one and ask whether a normal person could see a way out from its first screen. The Ancient Egypt level stays as it is for now, since the detour is real and short. But I'm going to keep an eye on its finish rate, because 49% says something the article titles never could.