The bailout rate is the percentage of users who leave a page before it loads and start looking for a faster, more engaging site. In their first "Need for Speed" study of 1999, Zona Research found that pages larger than 40K had bailout rates of 30 percent.  Once the designer reduced the same page to 34K, the bailout rate fell to between 6 and 8 percent, a dramatic decrease for just a few kilobytes. When fat pages were reduced to the recommended maximum of 34K, readership went up 25 percent.  These are averages, and users with faster connections and processors will experience faster downloads, but they can also become frustrated.
Zona's second study, "Need for Speed II," took into account dynamic transactions in order to modify the so-called "8-second rule."  They recommend that designers of dynamic sites cut an additional 0.5 to 1.5 seconds off connection latency in order to stay at the same level of abandonment compared with static web pages. As the web moves from a "plumbing" (pipes delivering pages) to a "transaction" (a series of dynamically generated pages) model, they argue that "cumulative frustration" plays an important role in user satisfaction.
Users can change the way they browse a site as they request and view additional pages. As they become more proficient, their learning "spills over," and users reduce their expected number of page views on returning visits . Clickstream-based analysis suggests that visitors trade off the number of pages requested and the time spent at each page.  Users may set "time budgets" for particular tasks, even though the tasks may take mul tiple pages to complete.
Without effective feedback, users will wait only so long for your pages to load. For longer delays, you can extend the time that users are willing to wait with realistic feedback. Displaying your pages incrementally, a crude form of feedback, can extend user tolerance for delays.
Myers found that users prefer percent-done progress indicators for longer delays.  These linear progress bars lower stress by allowing experienced users to estimate completion times and plan more effectively. Such progress bars are commonly used in download managers.
Bickford found that with no feedback, half of his test subjects bailed out of applications after 8.5 seconds. Switching to a watch cursor delayed their departure to an average of 20 seconds. "An animated watch cursor was good for over a minute, and a progress bar would keep them waiting until the Second Coming." 
Dellaert and Kahn found that wait time negatively affects consumer evaluation of web sites, but that this effect could be mitigated by providing information about the delay.  Delay information reduces uncertainty between expected and actual delays. For longer delays, they found that countdown feedback, a form of percent-done indicator, was better than duration feedback.
They also found that delays before viewing pages are less frustrating than delays while viewing pages. In other words, any delay after a page has loadedfor example, a sluggish response while users are interacting with the pageis worse than delays before a page has loaded.
Response times below two seconds are ideal, but current bandwidths make this speed impractical , so we settle for 8 to 10 seconds. What does this mean for web page design?
Page size and complexity have a direct effect on page display speed. As you learned in the Introduction, the majority of current users are at 56Kbps or less. That trend will continue until at least 2004, with international users lagging behind until 2007. Table 1.2 shows the maximum allowable page size needed to meet three attention thresholds at different connection speeds (derived from Nielsen, Designing Web Usability , 2000).
|Bandwidth||1 Second||2 Seconds||10 Seconds|
|These figures assume 0.5-second latency.|
You can see that 34KB is about the limit of total page size to achieve the 10-second attention threshold at 56Kbps. Under 30KB would be an appropriate limit for 8.6 seconds at 56Kbps. This is total page size, which includes ads and graphics. Assuming a 10KB banner ad and some small graphics, your HTML should be at most around 20K.
Designers who violate the 30KB limit will pay for their largesse with lost sales, increased bailout rates, and increased bandwidth costs.
The Limits of Short- Term and Working Memory
Chunking is important to the working of short-term memory. For example, we do not easily retain the number 7545551212 but have much less difficulty with 7545551212. There are three chunks and we know where to expect boundaries. Our short-term memory is limited in the number of "chunks" of information it can hold. As we gain expertise with an activity, we tend to think more abstractly and acquire shortcuts, increasing our overall chunk size and thus increasing how much we can perceive and accomplish.
Shneiderman and others suggest that delays increase user frustration due to the limits of short-term and working memory. Depending on complexity, "people can rapidly recognize approximately" three to "seven 'chunks' of information at a time, and hold them in short-term memory for 15 to 30 seconds." 
George Miller's original (and entertaining) "The Magical Number Seven, Plus or Minus Two"  study was subsequently shown to be a maximum limit to short-term memory span for simple units of information along one dimension. Broadbent argued that the basic unit of memory is closer to three, where each chunk can perhaps hold three further chunks.  LeCompte showed that as word length or unfamiliarity increases , memory span decreases, and Miller's maximum memory span should really be three to cover 90 percent of the population.  Mandler said that the magic number is closer to five.  Some 44 years after Miller's original paper, Kareev found that the effect of capacity limitations of working memory forces people to rely on samples consisting of seven plus or minus two items, for simple binary variables . 
People make plans on what to do next , and form solutions to problems while waiting between tasks. Longer delays (11 to 15 seconds) tax the limits of short-term memory and frustrate users who cannot implement their plans, and errors increase. Extremely short response times also increase errors, but lower their cost, improving productivity and encouraging exploration. More complex tasks require the use of working memory, which slows the optimum response rate. Given their druthers, users prefer short response times to long ones.
So what have we learned from all this? Speed of response is certainly one factor in user satisfaction on the web. Consistency of response times is another. But some researchers say that modeling human behavior in real-time environments with fixed performance metrics (like response times below 10 seconds) is too simplified. What we need is a more holistic approach.