To keep this post manageable in length, I have only included a small subset of the illustrative examples and discussion. I have published a longer version of this post, with more examples (but the same intro and concluding section), on my personal site.
Last year, during the months of June and July, as my work for MIRI was wrapping up and I hadn't started my full-time job, I worked on the Wikipedia Views website, aimed at easier tabulation of the pageviews for multiple Wikipedia pages over several months and years. It relies on a statistics tool called stats.grok.se, created by Domas Mituzas, and maintained by Henrik.
One of the interesting things I noted as I tabulated pageviews for many different pages was that the pageview counts for many already popular pages were in decline. Pages of various kinds peaked at different historical points. For instance, colors have been in decline since early 2013. The world's most populous countries have been in decline since as far back as 2010!
Defining the problem
The first thing to be clear about is what these pageviews count and what they don't. The pageview measures are taken from stats.grok.se, which in turn uses the pagecounts-raw dump provided hourly by the Wikimedia Foundation's Analytics team, which in turn is obtained by processing raw user activity logs. The pagecounts-raw measure is flawed in two ways:
- It only counts pageviews on the main Wikipedia website and not pageviews on the mobile Wikipedia website or through Wikipedia Zero (a pared down version of the mobile site that some carriers offer at zero bandwidth costs to their customers, particularly in developing countries). To remedy these problems, a new dump called pagecounts-all-sites was introduced in September 2014. We simply don't have data for views of mobile domains or of Wikipedia Zero at the level of individual pages for before then. Moreover, stats.grok.se still uses pagecounts-raw (this was pointed to me in a mailing list message after I circulated an early version of the post).
- The pageview count includes views by bots. The official estimate is that about 15% of pageviews are due to bots. However, the percentage is likely higher for pages with fewer overall pageviews, because bots have a minimum crawling frequency. So every page might have at least 3 bot crawls a day, resulting in a minimum of 90 bot pageviews even if there are only a handful of human pageviews.
Therefore, the trends I discuss will refer to trends in total pageviews for the main Wikipedia website, including page requests by bots, but excluding visits to mobile domains. Note that visits from mobile devices to the main site will be included, but mobile devices are by default redirected to the mobile site.
How reliable are the metrics?
As noted above, the metrics are unreliable because of the bot problem and the issue of counting only non-mobile traffic. German Wikipedia user Atlasowa left a message on my talk page pointing me to an email thread suggesting that about 40% of pageviews may be bot-related, and discussing some interesting examples.
Relationship with the overall numbers
I'll show that for many pages of interest, the number of pageviews as measured above (non-mobile) has declined recently, with a clear decline from 2013 to 2014. What about the total?
We have overall numbers for non-mobile, mobile, and combined. The combined number has largely held steady, whereas the non-mobile number has declined and the mobile number has risen.
What we'll find is that the decline for most pages that have been around for a while is even sharper than the overall decline. One reason overall pageviews haven't declined so fast is the creation of new pages. To give an idea, non-mobile traffic dropped by about 1/3 from January 2013 to December 2014, but for many leading categories of pages, traffic dropped by about 1/2-2/3.
Why is this important? First reason: better context for understanding trends for individual pages
People's behavior on Wikipedia is a barometer of what they're interested in learning about. An analysis of trends in the views of pages can provide an important window into how people's curiosity, and the way they satisfy this curiosity, is evolving. To take an example, some people have proposed using Wikipedia pageview trends to predict flu outbreaks. I myself have tried to use relative Wikipedia pageview counts to gauge changing interests in many topics, ranging from visa categories to technology companies.
My initial interest in pageview numbers arose because I wanted to track my own influence as a Wikipedia content creator. In fact, that was my original motivation with creating Wikipedia Views. (You can see more information about my Wikipedia content contributions on my site page about Wikipedia).
Now, when doing this sort of analysis for individual pages, one needs to account for, and control for, overall trends in the views of Wikipedia pages that are occurring for reasons other than a change in people's intrinsic interest in the subject. Otherwise, we might falsely conclude from a pageview count decline that a topic is falling in popularity, whereas what's really happening is an overall decline in the use of (the non-mobile version of) Wikipedia to satisfy one's curiosity about the topic.
Why is this important? Second reason: a better understanding of the overall size and growth of the Internet.
Wikipedia has been relatively mature and has had the top spot as an information source for at least the last six years. Moreover, unlike almost all other top websites, Wikipedia doesn't try hard to market or optimize itself, so trends in it reflect a relatively untarnished view of how the Internet and the World Wide Web as a whole are growing, independent of deliberate efforts to manipulate and doctor metrics.
The case of colors
Let's look at Wikipedia pages on some of the most viewed colors (I've removed the 2015 and 2007 columns because we don't have the entirety of these years). Colors are interesting because the degree of human interest in colors in general, and in individual colors, is unlikely to change much in response to news or current events. So one would at least a priori expect colors to offer a perspective into Wikipedia trends with fewer external complicating factors. If we see a clear decline here, then that's strong evidence in favor of a genuine decline.
I've restricted attention to a small subset of the colors, that includes the most common ones but isn't comprehensive. But it should be enough to get a sense of the trends. And you can add in your own colors and check that the trends hold up.
|Page name||Pageviews in year 2014||Pageviews in year 2013||Pageviews in year 2012||Pageviews in year 2011||Pageviews in year 2010||Pageviews in year 2009||Pageviews in year 2008||Total||Percentage||Tags|
Since the decline appears to have happened between 2013 and 2014, let's examine the 24 months from January 2013 to December 2014:
|Month||Views of page Black||Views of page Blue||Views of page Brown||Views of page Green||Views of page Orange||Views of page Purple||Views of page Red||Views of page Violet||Views of page White||Views of page Yellow||Total||Percentage|
As we can see, the decline appears to have begun around March 2013 and then continued steadily till about June 2014, at which numbers stabilized to their lower levels.
A few sanity checks on these numbers:
- The trends appear to be similar for different colors, with the notable difference that the proportional drop was higher for the more viewed color pages. Thus, for instance, black and blue saw declines from 129K and 126K to 30K and 41K respectively (factors of four and three respectively) from January 2013 to December 2014. Orange and yellow, on the other hand, dropped by factors of close to two. The only color that didn't drop significantly was red (it dropped from 84K to 67K, as opposed to factors of two or more for other colors), but this seems to have been partly due to an unusually large amount of traffic in the end of 2014. The trend even for red seems to suggest a drop similar to that for orange.
- The overall proportion of views for different colors comports with our overall knowledge of people's color preferences: blue is overall a favorite color, and this is reflected in its getting the top spot with respect to pageviews.
- The pageview decline followed a relatively steady trend, with the exception of some unusual seasonal fluctuation (including an increase in October and November 2013).
One might imagine that this is due to people shifting attention from the English-language Wikipedia to other language Wikipedias, but most of the other major language Wikipedias saw a similar decline at a similar time. More details are in my longer version of this post on my personal site.
Geography: continents and subcontinents, countries, and cities
Here are the views of some of the world's most populated countries between 2008 and 2014, showing that the peak happened as far back as 2010:
|Page name||Pageviews in year 2014||Pageviews in year 2013||Pageviews in year 2012||Pageviews in year 2011||Pageviews in year 2010||Pageviews in year 2009||Pageviews in year 2008||Total||Percentage||Tags|
Of these countries, China, India and the United States are the most notable. China is the world's most populous. India has the largest population with some minimal English knowledge and legally (largely) unfettered Internet access to Wikipedia, while the United States has the largest population with quality Internet connectivity and good English knowledge. Moreover, in China and India, Internet use and access have been growing considerably in the last few years, whereas it has been relatively stable in the United States.
It is interesting that the year with the maximum total pageview count was as far back as 2010. In fact, 2010 was so significantly better than the other years that the numbers beg for an explanation. I don't have one, but even excluding 2010, we see a declining trend: gradual growth from 2008 to 2011, and then a symmetrically gradual decline. Both the growth trend and the decline trend are quite similar across countries.
We see a similar trend for continents and subcontinents, with the peak occurring in 2010. In contrast, the smaller counterparts, such as cities, peaked in 2013, similarly to colors, and the drop, though somewhat less steep than with colors, has been quite significant. For instance, a list for Indian cities shows that the total pageviews for these Indian cities declined from about 20 million in 2013 (after steady growth in the preceding years) to about 13 million in 2014.
Some niche topics where pageviews haven't declined
So far, we've looked at topics where pageviews have been declining since at least 2013, and some that peaked as far back as 2010. There are, however, many relatively niche topics where the number of pageviews has stayed roughly constant. But this stability itself is a sign of decay, because other metrics suggest that the topics have experienced tremendous growth in interest. In fact, the stability is even less impressive when we notice that it's a result of a cancellation between slight declines in views of established pages in the genre, and traffic going to new pages.
The data for philanthropic foundations demonstrates a fairly slow and steady growth (about 5% a year), partly due to the creation of new pages. This 5% hides a lot of variation between individual pages:
The dominant hypothesis: shift from non-mobile to mobile Wikipedia use
The dominant hypothesis is that pageviews have simply migrated from non-mobile to mobile. This is most closely borne by the overall data: total pageviews have remained roughly constant, and the decline in total non-mobile pageviews has been roughly canceled by growth in mobile pageviews. However, the evidence for this substitution doesn't exist at the level of individual pages because we don't have pageview data for the mobile domain before September 2014, and much of the decline occurred between March 2013 and June 2014.
What would it mean if there were an approximate one-on-one substitution from non-mobile to mobile for the page types discussed above? For instance, non-mobile traffic to colors dropped to somewhere between 1/3 and 1/2 of their original traffic level between January 2013 and December 2014. This would mean that somewhere between 1/2 and 2/3 of the original non-mobile traffic to colors has shifted to mobile devices. This theory should be at least partly falsifiable: if the sum of traffic to non-mobile and mobile platforms today for colors is less than non-mobile-only traffic in January 2013, then clearly substitution is only part of the story.
Although the data is available, it's not currently in an easily computable form, and I don't currently have the time and energy to extract it. I'll update this once the data on all pageviews since September 2014 is available on stats.grok.se or a similar platform.
The following are some other hypotheses for the pageview decline:
- Google's Knowledge Graph: This is the hypothesis raised in Wikipediocracy, the Daily Dot, and the Register. The Knowledge Graph was introduced in 2012. Through 2013, Google rolled out snippets (called Knowledge Cards and Knowledge Panels) based on the Knowledge Graph in its search results. So if, for instance, you only wanted the birth date and nationality of a musician, Googling would show you that information right in the search results and you wouldn't need to click through to the Wikipedia page. I suspect that the Knowledge Graph played some role in the decline for colors seen between March 2013 and June 2014. On the other hand, many of the pages that saw a decline don't have any search snippets based on the Knowledge Graph, and therefore the decline for those pages cannot be explained this way.
- Other means of accessing Wikipedia's knowledge that don't involve viewing it directly: For instance, Apple's Siri tool uses data from Wikipedia, and people making queries to this tool may get information from Wikipedia without hitting the encyclopedia. The usage of such tools has increased greatly starting in late 2012. Siri itself was released with the third generation iPad in September 2012 and became part of the iPhone released the next month. Since then, it has shipped with all of Apple's mobile devices and tablets.
- Substitution away from Wikipedia to other pages that are becoming more search-optimized and growing in number: For many topics, Wikipedia may have been clearly the best information source a few years back (as judged by Google), but the growth of niche information sources, as well as better search methods, have displaced it from its undisputed leadership position. I think there's a lot of truth to this, but it's hard to quantify.
- Substitution away from coarser, broader pages to finer, narrower pages within Wikipedia: While this cannot directly explain an overall decline in pageviews, it can explain a decline in pageviews for particular kinds of pages. Indeed, I suspect that this is partly what's going on with the early decline of pageviews (e.g., the decline in pageviews of countries and continents starting around 2010, as people go directly to specialized articles related to the particular aspects of those countries or continents they are interested in).
- Substitution to Internet use in other languages: This hypothesis doesn't seem borne out by the simultaneous decline in pageviews for the English, French, and Spanish Wikipedia, as documented for the color pages.
It's still a mystery
I'd like to close by noting that the pageview decline is still very much a mystery as far as I am concerned. I hope I've convinced you that (a) the mystery is genuine, (b) it's important, and (c) although the shift to mobile is probably the most likely explanation, we don't yet have clear evidence. I'm interested in hearing whether people have alternative explanations, and/or whether they have more compelling arguments for some of the explanations proffered here.
So, the Android app doesn't report back page views, did I get that right? With tablets the gap between mobile and non-mobile is shirking, I use a tablet for reading the web very often even when having access to a laptop, I would not use a phone for that, and with a keyboard and perhaps a larger tablet I could do much of my work on it. (Fun anecdote: there is such a thing as a country with most of its current Constitution assembled on an iPad.) I don't want to take any position on the issue but I would recommend urgently fixing the Android / Apple app to report back page views, from my angle Wikipedia-in-a-browser is rapidly becoming obsolete. It is an encyclopedia. Encyclopedias are books. I read books comfortably on the couch, on a tablet or Kindle, since they don't require high-bandwith text entry. I don't go to the desk and crouch over the laptop to read an encyclopedia, it is bad enough to do it when working or writing. (I find it funny how Gizmodo asked if people still use tablets? People still breathe? To me, tablets ergonomically book-ifying my web reading experience was the best thing that happened since the invention of the wifi.)
Could you convert the tables into graphs, please? It's much harder to see trends in lists of numbers.
Another possible hypothesis could be satiation. When I first read wikipedia, it dragged me into hours long recursive article reading. Over time I've read more and more of the articles I find interesting, so any given article links to fewer unread interesting articles. Wikipedia has essentially developed a herd immunity against me. Maybe that pattern holds over the population, with the dwindling infectiousness overcoming new readers?
On second thought, I'm not sure that works at all. I guess you could check the historical probability of following to another article.
Regression to the mean is a potential problem when you choose to examine the most extreme data points in a data set (highly viewed wikipedia pages in this case).
I didn't pick them as points that were most extreme as of earlier years, I picked them as generically popular topics. There should be no particular temporal directionality to view counts for such pages.
Yes it can. Knowledge graph leads to less traffic at pages that have snippets, leads to less people clicking on to pages that don't.
Because even the simplest query for someone's birth date, the kind that the knowledge graph snippet can intercept, will sometimes lead into a night-long wiki binge.
Mobile is a larger platform than desktop 2015. That fact and the knowledge graph seem like very plausible explanations.
Eh? Desktop is still more than half:
I meant in general. I did not look up the numbers for Wikipedia.
The decline happened as a result of my indefinite banishment from Wikipedia. How many page views did I generate when I was active on Wikipedia? A lot more than I generate now that I'm banned!
I'm kinda kidding around but there's more than a kernel of truth in there. When Wikipedia was first created... there were more than a gazillion bits of knowledge missing. Over time though... the "easiest" bits were filled in. As all the lowest hanging fruit was picked... there were less and less people tall enough to reach the higher fruit. Clearly this resulted in a significant decrease in editing activity and by extension... a decrease in page views.
What percentage of the total decline in page views does this explanation actually account for? Beats me. It has to account for some though.
On a tangentially related note... a few weeks after famous economists die... I like to try and grab a screenshot of the page views for their Wikipedia entries. Their page views have a huge spike as their life/death is widely discussed... but then the page views decline pretty quickly afterwards. Unfortunately, in too many cases I've forgotten to grab screenshots. And the graph doesn't look as good when you have to go back in time. :( Why economists? Well... I think they'd appreciate it more than most famous dead people. Plus, it could be interesting/informative to compare their graphs in order to try and discern some useful information about something... economical.
In exchange for my explanation... how about you try and resolve Satt's Paradox? Or, you can try and predict if/when/how quarters up are going to replace thumbs up.
This would rely on a large fraction of pageviews being from Wikipedia editors. That seems unlikely. Got any data for that?
No data, like I said...
I did find this...
That confirms a decline in editors... and by extension... a decline in edits/pageviews... but no idea what fraction of the total pageviews decline it represents. It's probably pretty small.
The Google explanation probably represents a much higher fraction. For a while Wikipedia seemed to frequently be at the top of numerous search results. This would of course equate to considerable pageviews. Now it seems like Wikipedia results aren't as frequently as high as they used to be.