Search keyword analysis is a unique and powerful tool for the web analyst; it’s the only time users explicitly tell us what problem they are looking to solve by visiting our site. Elsewhere we divine user needs based on abstractions: aggregated pathing, counting button clicks, watching trends change from a baseline.
Even better, when a user doesn’t find what they’re looking for they refine their search term to be more specific. They’re not only tell us what they want, they help us give it to them.
Knowing what Safari users are looking for gives us a unique perspective on what content is successful and which technology trends are rising and falling. It also shows us where we’re failing.
As an example, I grabbed the top 200 search terms Safari received in the first week in November and plotted them by the number of pages that are read by a user immediately after performing a search (i.e. we gave the user what they were looking for) on the x-axis, and the percentage of users who immediately leave the site after performing their search (i.e. we didn’t give them what they were looking for) on the y-axis. The size of the bubbles is the number of individual searches each term received.
It’s the outliers that are interesting; in an ideal world all search terms would crowd in the bottom right with nice, big circles (like ‘python’, ‘oracle’). We need to improve our content for terms found in the top left, where a larger proportion of users immediately leave without reading anything; see ‘solr’, ‘puppet’, ‘json’ and ‘effective java’ (well, maybe that last search term is a bit of a holy grail…)
The search terms in the bottom left show subjects have both a low exit rate and have low engagement; these are broader terms, which are probably the starting point for deeper searches. One could argue that there’s a trend from generic in the bottom left to more specific in the top right. Users searching for ‘for dummies’ are probably at the beginning of their information quest; those searching for ‘solr’ either find what they are specifically looking for immediately (so no extended reading) or leave (high exit rate).
Obviously, the larger the data set (a larger time scale), the more insight we’ll get into our user’s needs.