Increase your website sales and conversions.
Website optimization tools that simply work.

Predicting Conversions on a Website using Visitor Behavior

In this case study, using behavioral metrics of a visitor, we will try to predict how likely is he to contact Wingify for more information. That is, can we predict in real-time, what is the probability of a visitor on the Wingify website to contact us for more information.

Introduction

Recently, Wingify (ContextSense API) got covered on ReadWriteWeb and Pluggd.in. Due to two mentions on such popular blogs, a lot of visitors arrived on the website to play with ContextSense. A majority of them went away after interacting for a while. Some tweeted about the product and a few actually contacted. In this study, we will see if we can predict:

  • Of all visitors, can behavioral attributes (such as number of pages browsed, time spent on site, hour of day of visit, referral source, etc.) successfully predict if a visitor will contact for more information on the API?
    • What makes visitors not likely to contact? How can we target those visitors who are not likely to contact (converted)?
  • Can we find any attributes which make a visitor interested in contacting but eventually does not contact?
    • How can we target those visitors (like for example offering discounts) to make them actually contact?

Data on Visitor Behavior

To kickstart the study, we would need data on visitors to analyze. We use our own product on www.wingify.com so, let us start by collecting data from Wingify analytics. Following pieces of information on (each of the) website visitors would be used to do the analysis:

  • Time spent
  • Number of pages viewed
  • Number of visits (sessions)
  • Time spent on landing page (ContextSense)
  • Whether visited homepage
  • Last page viewed before clicking on contact us
  • Page viewed immediately after visiting landing page (ContextSense)

Visitor Segmentation and Data Extraction

Next, we would create three visitor segments which are of importance to this study

  1. Those who visited contact us page and then contacted through contact us form (contacted_through_form)
  2. Those who visited contact us page and then did NOT contact through the form (almost_contacted)
  3. Those who did not visit contact us page (did_not_contact)

Creating segments is very simple with Wingify, as an example, did_not_contact segment would be represented as:

Now, let us extract all the data we require separately for these segments.

Data Analysis

Over a period of 5 days, 1500 visitors visited the website (all mainly due to the articles on ContextSense). The breakup of visitors into segments and data on them is shown in following table:

did_not_contact almost_contacted contacted_through_form
Number of Visitors 1345 137 14
Bounce Rate 66.84% 4.38% 0%
Time Spent on website (seconds) 129.63 472.4 1120.29
Page Views on website 1.7 5.82 7.21
Visited Homepage? 18% 36% 36%
Time spent on landing page (seconds) 79.68 209.05 761.86
Number of distinct visits (sessions) 1.16 1.45 2.5


There is a lot of interesting data here, let us analyze it point by point:

  • 10% (~150/1500) of the all visitors visited Contact Us page. Hence, a sizeable number of people are interested in exploring the contact information.
  • After visiting the Contact Us page, further 10% (~ 15/150) actually fill the contact us form and hit the Send the Message button. That is, the total conversion rate of the Wingify.com is 10% x 10% = 1%, which isn't fantastic. We need to see why people aren't contacting when they arrive on the contact us page. Were they looking for something else? Maybe a physical address and a phone number can increase the level of comfort? A perfect case for split test.
  • The ReadWriteWeb article had a direct link to Wingify's contact us page at the end of the article. Some of the visitors arrived on the website directly from that link. This explains why almost_contacted segment has a bounce rate of about 4%.
  • There is a clear differentiation in behavior of the three segments when it comes to time spent on site, number of page views and number of sessions on the site:
    • The visitors who did not contact spent an average of 2 minutes (120 seconds) on the site, browsed through about 1.7 pages and visited the site 1.16 times. Hence, we can be comfortable in saying that a new visitor who spends more than (say) 3 minutes and browses more than 2 pages on the site is a bit more interested in us than an average visitor. And if the visitor is a repeat visitor, then the confidence increases even further
    • As you can see in the table, there is a clear distinction also between visitors who had almost contacted and visitors who have actually contacted. contacted_through_form visitors spent almost 20 minutes on the site which is 3x the time spent by almost_contacted visitors. Perhaps the latter segment visitor did not see all the features of the product or missed some of the most interesting parts. This is confirmed by time spent by them on landing page (ContextSense). Visitors who actually contacted spent about 10 minutes on the landing page alone (though may be in multiple sessions). They played with ContextSense for a good amount of time. In contrast, the time on landing page for almost_contacted visitors was 3 minutes. The same story is repeated by other metrics: page views on site and number of sessions.
  • This data reveals that this situation is a perfect case for behavioral targeting. On the Contact Us page, we can target all those visitors who spend less than say 6 minutes on the site or see less than 7 pages on the site with content (presentations, case studies, whitepapers, product tour, etc.) to fully convince them that it is worthwhile to contact Wingify.
  • For visitors who have already spent a lot of time on site, we should simply let them contact us because they are convinced of contacting.

Funnel Analysis

For visualization and also to confirm our calculations, let us see a funnel (path) which we think visitors would have followed during their visit:

  1. Visited ContextSense page
  2. After browsing around the website, visited Contact Us page
  3. Finally, contacted by filling contact information in the form and hit send button


Interesting insights:

  1. As already noted, the total conversion rate is close to 1%
  2. In almost_contacted segment, some of the visitors (86-76 = 10) directly visited Contact Us page (perhaps from ReadWriteWeb article). So, these visitors are not part of last stage of this funnel
  3. In contacted_through segment, (12-10 = 2) visitors directly visited the contact us page

Prediction Model

Now we analyze actual raw data and use it to build a prediction model for conversions (in this case, contacts). Let's first export the data from Wingify in MS Excel compatible format:

We have exported all the data available for all 3 segments. Click here to download the data file which was directly exported from Wingify. The way to visualize the data is that the column containing segment name is the output of the prediction model and all other columns are inputs.

First, let’s try to predict if a visitor will visit the Contact Us page or not. Hence, we have two possible outputs that we need to predict:

  • Output value = 0. Those who did not visit Contact Us page (did_not_contact)
  • Output value = 1. Those who visited Contact Us page (almost_contacted and contacted_through_form)

Before processing it is also necessary to remove variables which we presume to be either entirely irrelevant or are not very relevant to the output. These are variables such variables are visitor id, last updated, browser, etc. For this, the raw data is processed to produce an excel which has just numbers.

Next, let us calculate correlations of different inputs (metrics) with the output (visited Contact Us page). The correlation values can be seen as the amount of influence a metric has on visitor's likelihood of visiting Contact Us page. (Statistically insignificant correlations at p = 0.1 are not shown )

The data above has interesting nuggets of information. The biggest insight is that the number of page views on Wingify.com has a tremendous impact on whether a visitor will visit Contact Us page or not. What is surprising is that the number of page views has almost 3 times the correlation value than the time spent on website. So, the learning here is that the website design and structure should be conducive to exploration of different parts of the website; overall time spent on pages is not as important as different number of pages seen.

Now let us do the same exercise again, but this time we will predict what influences visitor's likelihood to contact, once he is on the Contact Us page. Hence, our output values now change to:

  • Output value = 0. Those who did not contact after visiting Contact Us page (almost_contacted)
  • Output value = 1. Those who contacted (contacted_through_form)

Aha, surprise surprise! Number of page views has no bearing on whether a visitor will actually contact. Instead, what matters is the time spent on the landing (ContextSense) page. This is very important for various reasons:

  • It shows that the most important indicator of a conversion is visitor intent. Here, ContextSense captured the intent of the visitors and they played around with the demo for a significant amount of time before contacting.
  • It clearly displays potential for targeting. What can be done is that on the Contact Us page, we can target more information on ContextSense (examples, use-case scenaiors, API details, etc.) to the visitors whose landing page is ContextSense and have spent less than (say) 5 minutes on ContextSense page

Regression

To calculate numerically what effect does Time Spent on ContextSense page in minutes (X) has on Probability of Actually Contacting (Y), let us do a linear regression of the form Y = m*X + c.

Performing regression on the data we find (note that the following result is statistically significant):

Probability of Actually Contacting = 0.0176 * (Time spent on ContextSense in minutes) + 0.0417

Observe that the base probability of contacting is around 4.2% and every minute spent by the visitor on ContextSense page increases the probability of contacting by 1.8%. This, of course, assumes that the interest of a visitor increases linearly with time (when in fact it may be exponential).

We can use the above model to derive following actionable information. If we need to increase the current 10% conversion on Contact Us page to 25%, we need to make sure that the visitor spends at least (0.25 - 0.0417)/0.0176 minutes on ContextSense. The value turns out to be 12 minutes for this case.

While a value of 12 minutes on a page may look large, it is due to large time taken by ContextSense for doing the analysis, amount of time required by visitor to study its results, then trying out ContextSense with different URLs and finally trying out ContextSense across multiple sessions.

Summary

  • The total conversion rate for Wingify.com – for a recent wave of visitors from ReadWriteWeb and Pluggd.in – turned out to be 1%
    • Of all visitors arrived, 10% visited the Contact Us page (almost_contacted)
    • Out of those visitors who visited Contact Us page, 10% actually contacted (contacted_through_form)
  • As expected, there is clear distinction between different types of visitors as far as time spent, page views, number of sessions is concerned:
    • contacted_through_form > almost_contacted >> did_not_contact
  • The most important predictor for detecting if a visitor will visit Contact Us page turned out to be number of pages viewed by him. So, to get visitors interested enough to visit the Contact Us page, it is vital to have an engaging website where visitors can fully explore what we have to offer
  • For actual conversions (people who fill the contact us form and get in touch with Wingify for more information), the time spent on landing page (in this case ContextSense) turned out to be most important predictor. This shows that the landing page must match with the intent of the visitor
  • Ideal amount of time we would want a visitor to spend on the landing page before visiting the Contact Us page was predicted to be 12 minutes. Time spent on landing page is a proxy to how familiar/engaged a visitor is to what you are offering. In order to leverage this information, when the visitor visits Contact Us page, if we detect that s/he has spent less amount of time on landing page, we can persuade him/her to interact a bit more with the application, trying out different options and with different inputs. We can setup a split test to see if it actually increases conversions




← Back to Resources