For the past several months, the Portal team within Wikimedia Foundation’s Discovery department has been working hard on making visual and practical improvements to the Wikipeda.org Portal page to improve user experience. These changes include a search box with article suggestions, descriptions of other Wikimedia projects, and detecting language and dynamically populating/resorting the links around the globe. Looking back at various engagement metrics, we found slightly increases but engagement with the page has stayed largely the same. Of course, engagement does not correlate with satisfaction, and we currently do not have any methods to measure Portal visitors’ satisfaction and whether or not they feel welcome and end up visiting the page again or regularly.
{ RMarkdown Source | Figures | Analysis Codebase | Download PDF }
For the past several months, Discovery’s Portal UX team – Deb Tankersley, Jan Drewniak, Moiz Syed (formerly), and Julien Girault (formerly) – have worked hard to make visual and practical improvements to the Wikipedia.org Portal page. The goal was to improve the experience of wikipedia.org visitors and help them get to the information they are looking for as fast as porrible. These improvements include (but are not limited to) the following major deployments:
In Figure 1a, we see that by default the language links (“primary links”) around the Wikipedia globe logo are the top 10 languages by number of articles. The first language – English – is the default language of the search box that:
In Figure 1b, we see what visiting and searching looks like for a user who has set Icelandic as their preferred language, followed by Russian and Korean.
Figure 2 shows the links to other Wikimedia Foundation projects such as Commons and Wiktionary. The links are called “sister links”. As of 18 May 2016, they include descriptions of the projects. Currently, these descriptions are in English, but work is being done to translate them into the other languages so they dynamically show up in the user’s preferred language like the phrase “The Free Encyclopedia” does. See T142582 for more details.
We used the Siteviews tool on Wikimedia Tool Labs to obtain the pageviews for English, German, and French Wikipedias, which we used as the noncovariate in the Bayesian structural time series. Wikipedia.org Portal pageviews and clickthroughs were acquired from our Webrequest data (via this R script) and our Event Logging (EL) data (via this R script).
Wikipedia.org Portal EL uses this schema and a sampling rate of 1 in 200. Once a user visits wikipedia.org, they are randomly selected to be anonymously tracked with EL, we record:
Sessions last at least 15 minutes. The timer starts counting down from the user’s initial visit to the page. Any clicks they perform while that timer is counting down are recorded by our EL system. If they do not visit the page again in those 15 minutes, the next time they come back to the page we do not record any of their actions and remove the old tracker. After their session expires, they have the same 1 in 200 probability of being selected again. If they do visit the page again in those 15 minutes, we renew the expiration time onthe timer. For this reason, we may have sessions that last hours and have dozens (and even hundreds!) of visits and clicks.
Where applicable, we used generalized additive model (GAM) spline-based smoothing to fit a smooth curve to the data to visualize overall, longer-term patterns. Specifically, we used the s()
function in the mgcv R package, with a 4- and 5-dimensional basis for the smooth term (y ~ s(x, k = 4)
& y ~ s(x, k = 5)
) in majority of the graphs.
In Figures 8 and 9 below, we use the “most common section clicked per visit” metric. When the user goes to the Wikipedia.org Portal and gets enrolled into event logging by our random sampling mechanism, they have a tracker that stays alive for at least 15 minutes, with the expiration timer renewing every time the user visits the page within the 15 minute window. Furthermore, on each visit to the page, the visitor can click on any link as many times as they want and open as many tabs to different wikis as they want. This means that for any one page landing event (visit), we can see dozens (and even hundreds!) of clickthrough events, and that each session can consist of dozens (and even hundreds!) of visits. While the graph above uses the “last action taken” metric, the graphs below allow us to see how the most clicked section per visit varies over time.
We will be honest when we say that these are not the results we were hoping for. We were expecting to see each deployment followed by a dramatic increase in an appropriate metric – mostly engagement. However, these are also not the results we feared. It appears our efforts to make the Wikipedia.org Portal page look nicer and more welcoming have not really had an effect that can be measured through these metrics. We (currently) can’t measure overall satisfaction of the visitors. That is, if a person visits the page, feels more welcomed, and ends up going to the page sometime (specifically more than 15 minutes) later, we currently have no way of capturing that, and must rely on some of these engagement metrics.
On August 16th we deployed a design change that collapses the secondary links into a modal – see Figure 10 – which might or might not affect clickthroughs. In making it more modern looking and easier to see, we’re hoping (unmeasurable) user satisfaction will go up even if clickthroughs (engagement) might not exactly display that satisfaction. We will monitor the Portal analytics dashboard following the rollout.