Improving the ‘Stickiness’ of Your Website

Date: Sep 21, 2007

For a long time, the only solution to make websites appealing and "sticky" was to rely on gurus (web designers who were just supposed to know the "right" answers). But what if the guru made a mistake or did not take into account all the variables and created less-than-optimal pages? Alex Gofman explores ways to involve consumers in the co-creation process in the form of multivariate landing page optimization as a possible solution for the problem of the ever-increasing bounce rate on many websites.

It is hard to believe that one person alone could introduce so many revolutionary ideas: department stores, restaurants inside them, price tags, money-back guarantees, newspaper ads, white sales, commemorative postal stamps, etc., etc., etc.

All the above were invented by a man named John Wanamaker in the late nineteenth and early twentieth centuries. He reportedly coined the phrase "The customer is always right". And this amazing list can go on...

This merchandising and advertising genius is reputed to have complained that "Half the money I spend on advertising is wasted; the trouble is, I don’t know which half" (Zulker, 1993).

Visitors Make Snap Decisions

The world has changed since then, but the question remains. Virtual stores are replacing the brick-and-mortar ones. However, there is nothing virtual about inept advertising. Moreover, the problem is even more exacerbated if the advertisement actually works and brings the visitors to the site but for whatever reasons they leave without a purchase driving ROI down more and more.

A department store manager would be greatly alarmed if a big portion of prospective buyers walked in the store and in the first few seconds of their visit walked out of the door. This is exactly what happens to many websites—the visitors just "bounce back" from the landing page. "Bounce rate" of many sites is well above 50%.

Why does this happen? Did the visitor get there by a mistake? Or did he assume it was a mistake because he did not like the landing page or could not find the critical information fast?

Even worse, a visitor stayed on the site for some time, chose what she liked, but did not complete the purchase. According to MarketingSherpa data, the average ecommerce shopping cart has about 60% abandonment rate (Can you imagine three of every five department store carts left abandoned in the isles?). Is it because the visitors could not find some important information about shipping, taxes, return policy, etc.? Or is it because the site requires too much personal information?

Businesses understand the importance of making their websites appealing and "sticky." For a long time, the only solution was to rely on gurus: web designers who were just supposed to know the "right" answers.

What if the guru made a mistake or did not take into account all the variables? Indeed, people’s perception changes, the target visitors have their own unique preferences, etc. Potential loss of not optimizing the landing pages may be staggering.

As for the payment pages, many website designers do not consider that part of their sites important at all, from the appearance point of view. However, simple changes to those pages could bring a substantial improvement to revenue per visitor with some reporting boosting conversion rates as much as 300%.

A recent study by researchers in Canada showed that snap decisions that Internet users make about the quality of a web page have a lasting impact on their opinions. They also reported that impressions were made in the first 50 milliseconds of viewing.

The implication of these findings is that it is mostly the main features and the general appearance of the landing page that make a difference, not necessarily the actual content.

However, should we solely rely on artistic taste of web designers as the only solution? No one can replace a good designer. Luckily, there are some new ways to help them to achieve much better results faster.

In the last few years, an approach called Landing Page Optimization (LPO) became widely popular. The idea behind it is to create several prototypes and test them with consumers. In the simplest case of an A/B Split Test approach, there might be only two variations of a page.

On the other hand, the most advanced form of LPO, called Multivariate Landing Page Optimization (MVLPO), involves thousands and thousands of prototypes. Although MVLPO was developed in the late 1990s, it did not get the deserved attention until very recently, especially with the introduction of Google Website Optimizer.

A typical MVLPO creates multiple experimentally designed variations of a web page and evaluates the difference in the reaction or behavior of the consumers who visit these pages.

In one variation, a special Java script serves different executions of the page to the actual website visitors and traces their online behavior (e.g., Google Website Optimizer).

After accumulating sufficient data, the software finds the optimum combination of the elements of the page through a regression analysis. The method is quite reliable because it uses actual website visitors and monitors what they do, their conversion rate variations, and whether they stay or leave.

The problem with this approach is that it modifies the actual website, which could be considered risky by many operators. In addition, it assumes that some of the visitors will not like their page’s execution and then abandon the site. This may not be acceptable for businesses with low-traffic, high-stake websites and others.

An alternative solution uses experimentally designed web pages in a simulated environment typical for online market research activities. It is based on Rule Developing Experimentation (RDE) – a new paradigm developed in cooperation with Prof. J. Wind (Wharton Business School) and introduced in the book Selling Blue Elephants (Moskowitz, 2007).

Case Study

The operator of an online golf store wanted to optimize the landing page to increase the conversion rate and revenue per visit. As it catered to affluent golf players, the general traffic was not very heavy. However, the lifetime value (LTV) of customers was high because the site sold luxury and premium equipment.

The operator liked the idea of experimenting with the landing pages to improve their appeal, but she did not want to risk her biggest asset—loyal customers—to be disappointed by some potentially less-than-optimal prototypes during the experiment and leave. Instead, she decided to use MVLPO in a simulated environment using an RDE tool: Ideamap.NET.

The operator has several options for the banner, feature picture, and different promotions. What is the best combination of these components?

As you will see, with RDE, she found out to her surprise that it is quite easy and doable to answer this question even without much external help.

Collecting and Preparing Raw Materials

A landing page can have different layouts. In this case, the site contains a feature picture, a banner and three different types of promotions (see Figure 1). These placeholders are called silos or categories (Banner, Feature Picture, etc.).

Each of the five silos on the page has three options, called elements. There are many more possible designs (combinations of categories and elements) readily available for different layouts.

It is quite easy to test different layouts of the page with individual projects. The respondents in that case could be directed randomly to one of the projects.

The template is a schematic of the page; it places each element at a specific location. Readers familiar with Adobe Photoshop can compare the template with the layers in the graphical package. The process of a template creation is quite simple and involves positioning of sample elements (one from each category) on the page (Ideamap.NET).

RDE requires that sometimes an element be absent from the design (it allows for the estimation of absolute values of the utilities; for more information see Selling Blue Elephants).

In this case, the background of the template has some generic text and a neutral gray filling to be exposed instead of the missing options.

Figure 1 The template (in the middle, not to scale) and the tested elements of the website

Mixing and Matching the Elements According to an Experimental Design

The experimental design incorporated on the servers of the RDE tool automatically mixes and matches the elements using a template according to the underlying rules and presents them to the consumers.

The system dynamically creates a unique (up to statistically possible) set of 36 landing page executions for each respondent. The number of executions depends on the design (number of categories and elements).

Inviting Respondents and Collecting Their Ratings

There are multiple sources of consumers that can be invited to participate in the project. In many cases, visitors of the websites could be "intercepted" for a short survey.

In this case, the respondents were invited from a web panel. Six thousand invitations were e-mailed to random panel members. There were 431 responses, and 340 of them completed the survey.

Each respondent evaluated 36 unique pages out of thousands created by the experimental design (Figure 2 has two sample screens).

Respondents rated each of the screens on a 1 to 9 rating scale, answering this question: How interested would you be in purchasing GOLF GEAR from this website? (1=Not at all interested... 9=Very interested).

At the end of the survey, the respondent answered a short classification (demographic) questionnaire. One of the questions was used to screen out non-golf players. This process left 125 qualified respondents.

Figure 2 Two sample screens of the interview. Each respondent has a unique set of landing page variations.

Analyzing the Results

One of the key differentiating points of RDE is individual models of utilities for each respondent. Experimental design in RDE automatically creates unique balanced designs for each individual respondent. In a sense, these models allow deductions of the algebra of consumer minds. This helps to discover the patterns in the data, across elements and respondents, to generate rules for more targeted optimization.

Table 1 shows the utilities for the total panel (all the respondents) as well as for males and females. The utility presents a conditional probability of how many people out of 100 (percentage of people) will be interested in patronizing the store if the element is present. The higher the utility, the more interested the respondent would be in buying from this site. The negative utility means that the elements detract from the interest. The values around zero (+/-5 for this study) are generally neutral.

Table 1. Performance of the Elements for the Total Panel, Males and Females

		Total	Male	Female
Base Size		125	44	81
Constant		10	13	8
Banners
A3	Banner 3	0	5	-3
A1	Banner 1	-1	2	-3
A2	Banner 2	-1	6	-4
Promo A
B2	Free shipping	7	2	9
B3	$5.99 shipping	3	2	3
B1	Free $50 card	3	2	4
Visuals
C2	Golfer playing	16	17	16
C3	High-tech club	8	12	6
C1	Golf shoes	8	6	9
Promo B
D2	Final clearance—up to 65% off	12	11	13
D1	Save up to $100	8	8	8
D3	Free personalization	4	1	5
Promo C
E1	St. Andrews Sweepstakes	3	2	3
E2	115% price guarantee	3	3	3
E3	Golf vacation entry	0	0	1

The constant represents the basic interest of the respondents in this site if no elements are present. In this case, the constant is very low (+10), which means that in general the respondents are not very interested in a golf site (only 10% would be interested without seeing all the elements). What could change this perception?

The picture of a golfer alone on the golf course adds whopping (+16) points to the interest! In fact, all three feature pictures fared well with the golf shoes and the high-tech club scoring just one-half when compared with the picture of a golfer. This gives some valuable guidelines to the designer about what to feature on the site.

On the other hand, neither of the banners seems to play any role in changing the perception. Maybe the people just ignore the banners, knowing that there is not usually much useful information there. Or we did not create good elements and should go back to the drawing board.

The offer of free shipping generated some interest (+7), whereas sweepstakes and 115% guarantee did not. An average respondent might not feel that this has much relevance to a golf equipment website.

At the same time, the notice about the final clearance and savings up to $100 were quite impactful.

If we just pick the top-rated elements from each of the categories (see Figure 3), we can calculate the conditional probability of people buying from this site:

P = (Constant) + Sum (Utilities) = 10 + 0 + 7 + 16 + 12 + 3 = 48.

Figure 3 The highest-scored landing page for the total panel (the best combination of the elements)

This means that if we use the top elements from each category, we can increase the conditional probability of a visitor being inclined to purchase from the Web from 10% to 48%.

If we happen to choose wrong elements (the lowest scores in each category, see Figure 4), we will get the conditional probability of people buying from the site:

P = 10 – 1 + 3 + 8 + 4 + 0 = 24.

Figure 4 The lowest-scored landing page for the total panel (the worst combination of the elements)

This is exactly one-half of the optimal score. In other words, the price of a mistake could be one-half of your potential buyers!

The result of a simple permutation of quite similarly looking elements may have a very powerful impact on your conversion rate.

The size limitations of the article do not allow for the more in-depth RDE analysis of data, which usually includes segmentation, detecting synergies and suppressions between the elements, and so on. However, even top-line results (automatically produced and available in real time) are powerful enough for immediate actions with impactful results. I will discuss a more detailed approach in a separate article.

Discussion

Most of the projects that use respondents from online panels utilize a screener that selects respondents who are qualified to participate in the project. In our case, we screened unqualified respondents out during the data processing based on the classification questionnaire.

All the respondents who indicated that they do not play golf were removed from the dataset. This left us with 125 qualified respondents out of 340 that completed the survey.

The screening process makes data more targeted by removing the "noise" out of respondents’ data. Surprisingly, in this case, the removal of 63% of the total completed surveys did not change the data radically—the winning elements were winning and the losing elements were losing in both the full set and the screened subset (the absolute values of the utilities were different).

Specifically, the feature picture is similarly important to the full data set, to the players’ only subset and to any subgroup and segment we tested. One hypothesis is that there are some general rules of acceptability of landing page design for both users and non-users. This hypothesis, of course, has to be tested on many projects. If it proves to be true, it may have major implications for web design.

Conclusions

The John Wanamaker adage mentioned at the beginning of the article, "The customer is always right," has a much wider application than the usual customer relationship area. The actual customer could and should be involved in shaping everything from the design of an actual product to optimizing the marketing materials including websites.

Dr. Howard R. Moskowitz’s famous idea that people might not know what they want deep inside, but they will easily react if given the options, resonated widely though the media and industry (e.g., Malcolm Gladwell dedicated his entire speech on Ted 2004 conference to the implications of this idea).

To find a winner, one has to experiment and test multiple prototypes. Focus groups and A/B split tests employ simple "beauty contests" and cannot identify the real winning page design because they are limited to just a few executions.

On the other hand, MVLPO tests thousands of combinations and can find a real winner—for the total panel, subgroups, segments, etc.

Landing pages used to be an exclusive domain of web designers and highly compensated consultants, at least for the sites with large budgets.

For the rest, it was just the owners’ best guess. With the introduction of new tools, MVLPO made the field more democratic and available to virtually any website operator.

For the best results, the output of MVLPO could be fine-tuned by a designer who adds the final touches. Indeed, the designer could benefit tremendously by incorporating MVLPO in the design process.

Whether a website operator chooses to do it himself or with the help of a web designer who employs the approach as an integral part of the process; whether the experimentation is done on a live website with Google Website Optimizer or utilizes a more powerful RDE approach in a simulated environment, it will result in far better landing pages that in turn will increase the stickiness of the site and reduce the bounce rate.

Indeed, MVLPO is based on the quantification and amplification of customers’ voice and, as we know, the customer is always right.

References

Booth, Jeff (10/3/2006). Can Multivariate Tests Reduce Your Shopping Cart Abandons? Real-Life Results... Retrieved on 7/2/2007.
Gofman, A. (2006). Emergent Scenarios, Synergies, And Suppressions Uncovered Within Conjoint Analysis. Journal of Sensory Studies, 2006, 21(4): 373–414.
Lindgaard G., Fernandes G. J., Dudek C., and Brown, J. (March 2006). Attention web designers: You have 50 milliseconds to make a good first impression! Behaviour & Information Technology 25(2): 115–126.
Moskowitz, Howard R. and A. Gofman (2007). Selling Blue Elephants: How to make great products that people want BEFORE they even know they want them. Wharton School Publishing, 2007.
Theekson, Andy. Rocket Conversion Rates With Multivariate Testing. Retrieved on 7/2/2007.
Website Optimizer Overview (2007). Google. Retrieved on 7/2/2007.
Zulker, W. A. and J. Wanamaker (1993). King of Merchants: The Wanamaker Digest. Wayne, PA: Eaglecrest Press.