Earlier this week I stumbled upon a highly interesting research by Richard Gatarski, now assistant professor at School of Business, Stockholm University.

The research, actually over 5 years old is about evolutionary banner design. That is, using genetic algorithms to design advertisement banners in order to maximize their click-through rates.

The research was conducted on the e-commerce site boxman.se during an 11 day period in 1998. The basic layout of a simple banner design was broken into 4 areas and each area could be used to display various different elements about the product, such as the artist’s name, the title of the album, cover art, artist images and slogans. 5 different font types and 2 colors were also among the things that the genetic code could play with. (See more about genetic algorithm’s in yesterday’s entry.)

Using these features, an original set of 20 banners was randomly generated. The fitness of a banner was then measured by its click-through rate and the “fittest” designs used to breed the next generation, etc. For comparison, part of the banners displayed in the research were not a part of the evolutionary process, and hence remained the same during the entire period (for details refer to this paper).

The results were quite interesting. While the non-evolving banners had from day 1 a steady click-through rate of about 0.7%, the evolving banners grew over 9 generations from 0.95% to 1.41% (average for the entire generation). Same numbers, looking only at the numbers for the 5 most successful banners of each generation rose from 1.3% to 2.6% – doubling the click-throughs and more than tripling the rate for that stationary designs. Not bad!

Interestingly enough the last generation showed several trends, most notably that including the phrase “click here” plays a big role in the success of a banner. This is in line with earlier results from other academic research (oh, the simple creatures we are).

Given these results, I’m surprised that this practice doesn’t seem to be used for today’s advertisement banners, despite the obvious value. I also wonder if similar methodology could be used e.g. in usability design or even graphical design and the layout of web pages. The beauty of it being that users’ typical behavior is being used as a measurement of fitness, there is no additional survey or “Did you find this helpful?” links to click.

Details:
For a quick overview, you can browse through Gatarski’s presentation slides on the project from COTIM-99.

Details can be found in Gratarski’s papers from Generative Art ’98 and COTIM-99. The GA paper has more details on the design elements, while the COTIM one spends more time on methodology and results.