Chapter Four: Analysis of how users read hyperfiction
The aim of this chapter is to answer the question “How do users navigate a hyperfiction?” When confronted with a range of hyperlinks that break the sequential structure of a narrative, will the user capitalise on the opportunity for non-linearity, multiple perspectives and synchronicity that the mode of hyperfiction allows, or persist in reading in the established linear way? After outlining the successful construction of a hypertext in the previous chapter, it was important to investigate how the hypertext would then be read. This led to the decision to insert GA tags into the different hypertext pages, clearly marking the path taken by each user through the hypertext. The methodology involved in performing GA analysis is detailed in this chapter, as well as the challenges and difficulties in isolating individual users and paths. The expected results of the analysis are clearly stated that posit the different possible directions that the users could take through the hypertext. These results are then contrasted with the real data taken from the GA analysis that provide a clearer understanding of how users progress through a hyperfiction text.
Methodology: Applying Google Analytics
As detailed in the previous chapter, it was ultimately not possible to use the Twine hyperfiction together with GA. This resulted in the decision to publish to the web the individual sections of the hypertext as 28 separate HTML pages.
Employing GA in this way was extremely straight-forward. First it was required to set up a GA account, where the user receives a unique user ID. By setting up a new property, or web site, GA provides the user with a snippet of code containing the user ID that can be pasted into the heading of every webpage connected to the website.
I also inserted event tags, containing the name of the link and a user-assigned value, to each button or hyperlink. When the hyperlink is clicked on by the user, the information is recorded and stored by GA. This can be seen in Figure 9, in the HTML containing the tracking code at the head of the webpage, and event tracking code in every hyperlink. However in the end it was not necessary to use the event tags, as the information provided by the script tag in the head of a webpage was sufficient.
Figure 9: HTML containing GA tracking and event tags.
Once the GA tags were successfully inserted, the hyperfiction was complete and ready to be disseminated to readers. As stated in the previous chapter, access to the Twine hyperfiction community had to be forgone for the successful implementation of GA. To counterbalance this, the hypertext was promoted through social media, through my Twitter, WordPress and Facebook accounts.
Retweeted and shared on Facebook, it can be said that the hyperfiction trended for a very brief period. With GA it is not possible to access the total user data until the next day, when it supplies the average user session for the previous day. However it is possible to access the user information in real-time where the GA account-holder is given a window of use for the previous 30 minutes. At its highest point, the total number accessing the hyperfiction at once was five users, seen in Figure 10. This was shortly after the hyperfiction had first been promoted online, at the peak of its trending period.
It is interesting to note the high number of mobile users in the diagram, which actually outnumbers desktop users. This is a demographic that had never been considered before, which is significant because of the affordances of using mobile devices, and reading the hypertext on a smaller screen.
Figure 10: GA Real-time Analysis.
Overview of Google Analytics
From the start of the next day once the hypertext had been shared publicly, GA provided the average findings, included in Figure 11. The graph representing the number of total users surged on the day the hypertext was published, with 70 out of an eventual 87 users. The sharp drop-off could be as a result of the drop in social media coverage, or a lack of people sharing the hyperfiction or returning to read it.
There were a total of 1,574 page views, which divided by 87 (the number of users) gives an average of eighteen pages per user. This is quite encouraging considering there are 28 pages that make up the whole hypertext, as well as the large number of users that quickly drop off, mentioned shortly.
Figure 11: GA user overview.
Also of interest is the average length of sessions: 8.22 minutes. This is very high considering how much time people normally spend on a website, fifteen seconds according to a recent article on Time magazine (Haile). There are several sessions where the user spends near to that amount of time reading the hypertext, or where just two seconds on a webpage is clocked by a user (usually meaning they are clicking through to the next page), which serves to drive the average duration of a session down. But in the hypertext’s introduction I state that the length of time it should take to read the hypertext fully is 20 minutes, so the average session duration is lower than expected.
Extracting User Information
To analyse the hyperfiction user data, it was first necessary to extract the information relating to how the user navigated the hypertext. This proved quite complicated and time-consuming. While the information average was conveniently supplied by GA, the data for each user was not readily available. This has to do with the emphasis placed on anonymity and identity-protection by Google, which did not reveal a user’s IP address, and made arriving at the user’s path very indirect and long-winded.
Although it was not possible to view the path of every individual user, different sets of restrictions could be applied until an individual path was more-or-less reached, with their anonymity remaining intact. This process involved quite a long set of restrictions being applied, eg Country + Device + Browser Type + Screen Resolution + Screen Colour. This example may throw up two or more user sessions out of an original 130, where each individual path can be accessed with the User Flow facility, seen in Figure 12.
The User Flow model visualised the movements of the user through the website, diverging when the paths branched or split, and converging at nodes shared by other user paths farther along the flow diagram. This User Flow path was instrumental in helping to identify or define the different shapes that the user’s progress would take, thanks in part to its horizontal visualisation of the their paths. It was also very simple to use, it being possible to drag the chart along in order to progress, with the option of isolating specific paths proving very useful in identifying individual user sessions.
Figure 12: User Flow path highlighting individual user paths.
After the 130 user sessions had been whittled down, their interaction with the hypertext could then be examined. From this point the different users’ paths were recorded on a Microsoft Excel Spreadsheet. However as it was not possible to access the individual users (their protected IPs), the process of manually recording the separate information from different user sessions, then applying increasingly restrictive sets of conditions to determine which sessions were taken by whom, took place.
By applying increasingly specific restrictions, it was possible to identify two or more sessions made by the same user. For example, out of a total eighteen Dublin users on the Trinity College Dublin Internet network, eight were using the browser version Chrome 44.02403, seven were using the computer operating system Windows 7, seven had screens with resolution 1600×900 pixels, the same Flash and Java versions, and other restrictions.
Through passing these sets of conditions, the seven different user sessions can be taken to be by the same user. The possibility of then adding the data up from two separate sessions to make up one total user was briefly considered. However, as in the GA graph seen in Figure 11, only 33% of users were multi-users, and fewer still would have had their sessions included in the spreadsheet, as most multi-user sessions lasted only a few steps or web pages, so were removed. Also it made sense to analyse the user sessions rather than the users themselves, first in that it was physically not possible to analyse the users individually, aside from adding their sessions together, which would have been counter-productive; second because it was their sessions that were important, less so the individual user. It may be that with the same reading patterns carried through the repeated sessions, causing the pattern to be duplicated, but the sample size was not large enough for this to make a noticeable difference.
Figure 13: Excel Sheet recording paths of different users.
I was aware from the beginning that by including a brief introduction together with instructions, how I could be influencing the reading pattern of the user. For example, I describe the hypertext as “four parallel narratives” – stating this could cause the reader to adopt a reading pattern whereby they read the first four sections, and then the next four sections, in a zigzag or diagonal shape. Such an effect was unwanted because it would limit the findings of the experiment, which was to examine how users progress through a hyperfiction of their own accord.
Also in the dramatis personae page, the characters’ order of appearance is based on their car, together with a brief description stating the relationship between them. In this way, the son character and his father are positioned first and second, followed by the couple in the next car, the third and fourth characters. From this the reader may have sensed that an implicit order was in place, and so may have read the hyperfiction in this descending order. Due to the character’s name appearing first, I anticipated that the majority of the user sessions would consist largely of Lawrence’s narrative. I also expected that the narratives of Lawrence’s dad, and the characters Barry and Debra, would be read fewer times.
To avoid this and other features affecting the way that users read the hypertext, I tried to keep my instructions to a minimum. This was to force the user into partial uncertainty, allowing them adopt their own rubric for reading, and not to influence their reading in other ways. With the limit placed on instructions, I expected that the user would make several exploratory leaps using the hyperlinks, coming to grips with the hyperfiction structure, and then back-track to their initial starting point, to start again in earnest.
Regarding the user’s approach to the hypertext, I formulated three different directions or shapes that their progress might take. The first was an L-shape, where the user proceeds through a single narrative strand, and on reaching the end of that narrative, moves horizontally to read the final sections of the other parallel narratives.
I feared this to be the most common way of reading the hyperfiction, as it would be advancing in a linear way through each narrative that the user is normally accustomed to. This approach would mean opting out of the potential simultaneous narratives, in addition to multiple perspectives, with the user gaining a partial understanding of the overall story. It was logical to think that this approach of going from first-to-last and top-to-bottom, being the most common (some would say conditioned) way of reading, would dominate.
The L-shape style of reading would resemble this pattern in the data results:
Dad 01 -> Dad 02 -> Dad 03 -> Dad 04 -> Dad 05 ->
Dad 06 -> Barry 06 -> Debra 06 -> Lawrence 06 -> Dad 06
The next style of reading I expected was for readers to adopt a zigzag approach, moving from parallel sections on the same plane to the next section in this sequence, and across again, in a Z-shape. It is my belief this would have resulted in the most satisfying reading experience, with the user exploring the many different narrative techniques that the form of hyperfiction allows, such as interactivity, non-linearity synchronicity and multi-perspectives.
An explicit use of this kind of hyperlink can be seen the fourth section, when each character observes or adjusts the radio in some way, and the single-direction hyperlink transports the user in this spiralling, zigzag way.
The Z-shape would be represented in the data by this pattern:
Lawrence 01 -> Dad 01 -> Barry 01 -> Debra 01 -> Debra 02
-> Barry 02 -> Dad 02 -> Lawrence 02 -> Lawrence 03 ->
The final reading approach I expected to take place was in a U-shape, with the reader advancing first in a linear way to the end of a section, shifting to a parallel section, then backtracking to the start of that section to continue reading from there. This would have resulted in a more complete experience than the first shape or pattern, ultimately reading much more of the total hypertext. While this way does not constitute the total rejection of linear narrativity that the Z-shape represents, it does break the narrative mould far more than the L-shape. Although the user reads the text in a relatively linear way, by exercising the facility for re-reading the text, they can be said to adopt a non-liner, fragmented style of reading, where the features of multi-perspectives and synchronicity can still be observed.
Reading in a U-shape would take the following pattern:
Barry 01 -> Barry 02 -> Barry 03 -> Barry 04 -> Barry 05 ->
Debra 05 -> Debra 04 -> Debra 03 -> Debra 02 -> Debra 01
Google Analytics Results
In analysing the data it was first necessary to apply a set or conditions in order to remove results that could warp the findings. For instance, removing the large number of sessions lasting only two or three minutes reduced the number of user sessions considerably. Then the sessions of users were recorded, with each step taken represented by the web page bearing the character’s name that was accessed.
Removing user sessions that took too few steps or hyperlinks also occurred, as they held no value in being analysed. For example, if a user took less than four steps with the hyperfiction, not including the first four pages every user must pass through – the title page, introduction, instructions and dramatis personae – then it would be very difficult to conclude which of the different reading patterns that the user conformed to. Applying these different sets of criteria dropped the number of workable user sessions from 130 down to 50.
Of the different patterns or styles of reading, the Z-shape ranked highest, occurring 29 times overall in 50 total sessions. This demonstrates how a high number of the total users (58%) favoured a strictly non-linear approach to reading the hypertext. These users took advantage of the narrative affordances of hyperfiction, with its breaking-up of narrative linearity, plus its potential for multiple perspectives and synchronicity.
One explanation for the high number of users reading in this way, apart from exploiting the hyperlinks for a more interactive reading experience, may be the more general structure the pattern of the Z-shape has. While the zigzag pattern is clearly visible from the data, there is also quite a large amount of deviation within that: this style of reading rarely involves the careful study of each passage in a zigzag pattern, but a more approximate downward diagonal style of progression, with whole sections often completely skipped. This is unlike the other reading patterns outlined above, whose classifications are much stricter.
It is significant just how diverse the different paths of these Z-shape users can be, each of their paths containing great variety. This is a victory for the interactivity connected with hyperfiction, that none of these 29 users would have been met with the same experience while reading, they were all entirely unique.
The next most common style of reading the hyperfiction is in the U-shape pattern, occurring twelve times (24%). The U-shape represents a middle point between the non-linearity of the Z-shape, and the more rigidly linear L-shape pattern. While the user did proceed first in a linear way through the opening narrative, resisting all hyperlinks except the front button until the end, when faced with no alternative they returned to the start to read a parallel narrative from beginning again.
Finally nine users, or 18%, read in an entirely linear way, assuming the L-shape style of reading. When the U-shape is taken to be non-linear also, 82% of the total readers participated in a non-linear reading of the hypertext, which is encouraging. However for 18% of users to disregard the high number of hyperlinks that function to disrupt the narrative linearity, shows how ingrained the standard way of reading is.
Related to this was the decision to record how many of the user sessions featured the first character of Lawrence most prominently, especially towards the start of the session. Sixteen user sessions out of 50 were made up predominantly of readings of Lawrence’s narrative. This is quite high considering the next character to be read predominantly was Dad, with nine user sessions. The other characters were principally read even fewer times, Barry with four sessions, and Debra with two.
This does not add up to 50; it should be noted that not all of the sessions can be said to favour one character over another, and in many cases there is an even split between the narrators. But this result does represent a failure of the hyperfiction in breaking up the standard linearity that users select the very first named character, and do not deviate from that course of reading one character at a time from top to bottom. As expected, the order in which the characters appear may have influenced the direction that readers take.
Also a direct correlation was noted between the sessions that were linear and featured Lawrence most heavily: nine of the ten L-shaped user sessions were weighted towards the first character Lawrence, with the remaining seven user sessions pulled from the U-shape pattern of reading. This suggests that a large number of the users adopting a linear approach to reading the hypertext selected the first name in the list of characters, which is consistent with their preference for sequential reading.
While logging the user results in a spreadsheet, I recorded how many of the 50 readers utilised the radio station-hyperlink seen in section four of every character’s narrative. Occurring whenever a character observes or interacts with the car radio, this forms an example of multi-perspectives in addition to a potential hyperlink that connects the different characters together. Such use of the hyperlinks would be clearly visible in the findings, seen in variations on the following pattern:
Lawrence 04 -> Dad 04 -> Barry 04 -> Debra 04 -> Lawrence 04
However there were only nine reported uses of this hyperlink, making it into 18% of user sessions. This may seem low, but at that late point within the narrative (the fourth section), the user may have already become accustomed to clicking on hyperlinks located at the bottom of the webpage, rather than within the text itself. Also they may simply not have seen the blue hyperlink in the text. Nevertheless this way of hyperlinking formed a very useful experiment in permitting the reader some interactivity and autonomy, as well as allowing them to shift between the various different narratives simultaneously.
Lastly, the anticipated exploratory leaps taken by users at the start were not evident in the results, with it appearing that the users grasped how to use the hypertext right away, based on their reading patterns. From this it can be concluded that the average user is very familiar with hyperlinks and web pages, the hyperfiction not representing the alien concept that it was originally thought to be. This may not have been the case however with the large number of sessions that were cut short – perhaps those users, on being frustrated by the lack of guidelines and range of options, decided to end their session instead.
Some feedback by users has suggested the inclusion of a Back-button feature that will transport the user back to the list of characters at the end of each section, from where they could start to read a different character’s narrative. This would have been more convenient for the reader, saving them time in their overall session. Also as regards analysing the GA results, there would be less instances of the user backtracking through pages they have previously accessed, making the data clearer and more easily analysed.
However I did feel it was important to try and influence the decision of the reader as little as possible, and providing such a button may have unavoidably done so. Also the action of pressing the Back button six times in order to reach the character list once more did not seem overly arduous. When analysing the user session data, it was quite clear to see which users were actively reading the text and which were simply clicking through to another page, the average session length of the latter group being only two seconds on average.
On the writing of the actual hypertext, I would consider trying to make each section more independent. For example, a frequent criticism was that users after advancing through one character’s section, would then read the parallel section of another character’s, and be thrown into confusion. Such a hypertext as Jackson’s Patchwork Girl, made up of self-contained lexias which it’s possible to read in any order, had been toyed with in a previous chapter of this thesis, but was thought to be too time-consuming to execute, and unsuitable for parallel events, which this hypertext set out to describe.
Another possibility is to include more hyperlinks; this would have resulted in more sections or lexias than six per character, or else a longer hypertext than 6,200 words. I would also explore more levels of interactivity in the narrative structure, for example the possibility of alternative endings. This is such a staple of hypertext and interactive fiction that the decision not to include alternate endings seems unusual. However once again these branching endings did not seem appropriate to a parallel text.
59 out of 130 users read the hypertext on mobile devices, which is 45% of the total users. This high proportion of mobile users came as a complete surprise, and were I aware of this in advance I believe I would have worked to alter the structure of the hypertext. This may have involved making it more fragmented, with more sections containing fewer words in each, to make it more amenable to these users viewing the text on smaller screens.
The integration of GA tags with the hypertext enabled the querying of how users read hyperfiction, and produced very significant results. The different approaches to reading the hypertext were explored, and a taxonomy of different patterns of reading suggested. The most linear of these readings styles is the L-shape way of reading; reaching a midpoint between linear and non-linear is the U-shape reading pattern; lastly the Z-shape is put forward as a style of reading that the user adopts, this being the most linear and open way of reading the hyperfiction. The results from using GA found the Z-shape to be the most common reading pattern, adopted by over half of the users, who capitalized on the large number of hyperlinks within the text to disrupt the linear sequence. It was also found that almost one in five users maintained a linear reading pattern, revealing that when given the option of non-linearity, some users must still feel more comfortable assuming a linear way of reading text.
 Haile, Tony, “What You Think You Know About the Web is Wrong”, Time.