When designing web-pages for intra-net applications the notion of search time become important. In such applications, there are typically a limited set of web-pages which users will traverse frequently. For instance pages including contents lists and lists of links to reference documents. These pages will be scanned many times each day and to achieve efficiency the layout should be chosen to optimize average search time, rather than to optimize legibility for the occasional user. In this paper, we describe some results from a series of experiments on skilled users scanning of a screen display, that are relevant also to web-page design. The effect of variations in page layout features on the average search time was measured. It was found that with a fix page layout, learning takes place so that frequent users develop effective scanning strategies. These strategies are adjusted to the probability of finding interesting information in different locations on the page. It was also found that scanning a horizontal listing of items is slower than scanning a vertical listing of items. Findings further indicate that scanning a single long vertical list is faster than scanning multiple shorter vertical lists. Fixed position is the key to fast scanning. Users learn the location and directs search immediately to the correct location on the page. If the target item was given a unique feature, search time was, as expected, significantly decreased. There was no significant difference between the features: colour, shade, space, size and slant.
With the increasing development of intra-net applications many professionals are interacting with a web-browser during a substantial part of their working day. For this user category efficiency is essential. Because of this it is interesting to study how page layout features affects the average search time for a frequent user.
A classic model of search in monochrome alpha-numerical displays was constructed by Tullis, 1983 . He measured the search time for many different displays and used multiple regression to identify the layout characteristics that had a significant influence on search time. Although this modelling did not attempt to explain the process of visual interpretation, the insights gained have had a strong influence on practical display design.
Since Tullis there have been little empirical data reported on the effect of layout features on search time. One of the reasons for the lack of data in this area is that there are few studies of skilled performance. In most studies that compare design alternatives, a rather low number of trials have been used in the experiments. The results thus describe the behaviour of the novice, but not necessarily that of a trained user. When trained performance is assessed, the results may be completely different.
For instance, menu-search was studied by Somberg  Alphabetically arranged menu items were found to be searched significantly faster than a positionally constant arrangement of menu items. When Somberg extended the experimental series he found that after training the result was reversed. Evidently, training have some effect on visual search performance.
To come up with recommendations for display design for the frequent user, especially in a professional work environment, we need to understand exactly what happens when a user becomes skilled in the interpretation of a visual display. What is it that the user have learned? In what ways are the search strategies changed? How can we design display layouts to support the development of effective search strategies? To answer these questions we have performed a series of experiments where visual search performance after training have been assessed. The complete results are reported elsewhere , . The results that have a bearing on the design of web-pages will be described.
Our method of studying the scanning of a page is based on measurements of the search time; that is, the time required for searching out a target item on a display. After a number of trials where the position of the target item is varied in a controlled way, we can assess search time as a function of the position on the page. We can then reason about the strategy used for scanning the page. The output part of the interaction is kept constant; responses are always given by means of a mouse click.
In half of the trials the target item was present on the page. In the other half of the trials the target item was not present on the page. The correct response was dependent on the presence or absence of the target item. The answer was given by pressing one of two mouse buttons.
All experiments used basically two presentation conditions. In one condition, the layout supported only serial self-terminating search. In another condition, the layout was set up to support the development of a short cutting scanning strategy. The subjects were not informed about any differences between the conditions.
The subject was sitting in front of the computer screen with a viewing distance of 55 cm. A trial started with the presentation of a fixation cross in the middle of the screen for 200 ms. A screen image was then presented until the answer was given. A rest period of 400 ms followed before the cycle was repeated. Responses were logged and error rate calculated.
Search time was measured, for each trial, from the onset of the stimuli until the response was made.
A within-subjects design was used. The presentation condition served as the independent variable and the dependent variable was search time.
The experiments were controlled by an IBM RS-6000 320H computer equipped with a GTO double buffered graphical subsystem. The screen images were presented on an IBM 6091 19" (1280 x 1024) colour display.
All subjects in the experiments served voluntarily. They had computer experience of at least two years, and their vision was normal or corrected to normal. None had any known problems with colour vision.
The first blocks of data were cut out so that comparisons were based on data from the blocks of trials where no further improvement was evident.
The search process was modelled by a simple linear model. The best fit of the model parameters (including scanning rate) was determined by multiple regression. The models were validated by using the split-half method .
Results from the experiments indicated that:
The speed of scanning a page of familiar layout improved about 25% with practise, but became stationary after about 200 trials. This shows that results from studies with a low number of trials must be viewed with caution when generalized to skilled performance. Also, it shows that, for simple search tasks, it is possible to assess skilled performance in a laboratory setting without utilizing an unreasonably long training period.
The scanning rate (time per item) for close and equidistant vertically oriented items was estimated to about 100 ms; for separated items it was estimated to about 220 ms. The scanning rate of horizontally aligned items was estimated to be 1.2 times that of vertically aligned items. This shows that vertical alignment of data is preferable and that data items within a group should be grouped close together.
The estimated scanning rate was dependent of subject´s age. This is consistent with other findings, but the effect is small and hence unlikely to have any practical consequence.
The average time for comparison of adjacent numerical values was estimated to about 450 ms.
As expected, searching for a target item was significantly faster (83%) if the target item was given a unique feature, compared with if no unique feature was used. The colour, shade, space, slant and size features were equally effective compared with a control condition without features.
The figural pattern formed by varying number of digits in a list of numerical values was learned by the users. If such patterns were present, then this resulted in faster trend assessment than in a control condition where no figural patterns were present. This effect was more pronounced in a vertical orientation (42% faster), but existed also in a horizontal orientation (11% faster). This shows that alignment according to the decimal point is important. Also, an implication of this finding is that, in columns of words, the pattern of varying widths can probably be learned, and even improve orientation. Therefore, columns of words (e. g. menu options) should not be abbreviated to an equal number of letters.
If the probability of finding an interesting item was non-uniformly distributed across the screen area, then this pattern of varying probability was learned by the users. The search strategies were suitably adopted and this resulted in significantly faster search (32%) than if the probability distribution was uniform. A post experimental test showed that the subjects were not consciously aware of the probability distribution, but in spite of that, the search strategies were almost optimally adapted to the pattern of varying probability. This finding reveals that even if there are many items on the screen, the average search time can still be short if the probability of finding the target item is non-uniformly distributed across the screen area, and this distribution is constant over time.
The slowest search is obtained for multiple groups of horizontally aligned items in a random order.
The fastest search is obtained when search space can be restricted by knowledge captured by the peripheral visual system. Highlighting, informative patterns and spatial constant positions are such design principles. Presenting data items in fixed positions on the screen, so that each position has a meaning, is thus useful for effective interaction.
Grouping is not in itself an efficient design principle. It is faster to search within one large group of randomly ordered items than if the items are separated in groups. However, if the items are sorted and then grouped, search can be restricted to the target group only. Therefore, sorted groups, where each group has a fixed position on the screen, is an effective design principle.
These experiments have assessed factors that influence a skilled users scanning of a completely visible page. In future work we will address the issue of scanning a page where the whole content can not be viewed simultaneously. This means that navigation activities like scrolling commands have to be included in the search model.
1. Tullis, T. S. (1983). The formatting of alphanumeric displays: A review and analysis. Human Factors, 25(6), 657 - 682.
2. Somberg, B. L. (1987). A Comparison of rule-based and positionally constant arrangements of computer menu items. In Proceedings of Human Factors in Computing Systems, CHI ´87, ACM, 255 - 260.
3. Nygren, E, Allard, A. & Lind, M., Skilled Users Interpretation of Visual Displays. Submitted.
4. Nygren, E., Allard, A., Display design principles based on a Model of Visual Search. Submitted.
5. Murphy, K. R. & Davidshofer, C. O. (1988). Psychological Testing. Principles and Applications. Prentice-Hall International, 67 - 69.