Predictive Consumer Ad Targeting.

Theory on how to show the ads the visitor wants, before they arrive.

[killroy]

I've always been in love with split testing and statistics, and I'm currently fashioning a theory for even more powerfull ad targeting.

I have a site about Widgets (liek jsut about anybody on thisboard appearantly),which come in many colours. Lets, for the sake of simplicity asome I got red, green and blue widgets. So my site looks like this:

www.widgets-r-us.com/
www.widgets-r-us.com/red-widgets
www.widgets-r-us.com/green-widgets
www.widgets-r-us.com/blue-widgets

I have some live stats, that tell my at any time the order of the widgetpages ranked by impressions, giving me a moment to moment popularity countfor each colour. Sicne in reality I have many more then 3 colours, I can identify the long term popular ones, like the top 10, even thought theymight fluctuate up and down alittle each day.

Now each of those pages sells widgets. But I'Ve found that if I take the most popular widget, and place an ADDITIONAL site-wide banner for that one I can double the sales and make more profit. If I place a banner for the leasst popular widget, I make nothing in addition.

Now since my widget sales links are nicely organised, I can place any link I like at any time. So it would make sense to always place 1-5 links in the order of the current popularity.

The problem with that is that I base my decision always on how the popularity was over the LAST 24hours not the NEXT 24hours.

So my thought goes like this: Consider if I take a smoothed quadratic approximation of hour to hour (or 6hour to 6hour) ranking over the last few days. By simpletrend recognition and quadratic extrapolation, I should be able to predict where the raning will be in the next few hours, rather then jstu relying onwhere the trend was in the last few hours.

So I should be able to show the ads that the CURRENT VISITORS want to see, instead of the usual practice of showing the ads the PREVIOUS VISITORS want to see.

I think I'll mull a bit more over these ideas, but I'll make sure to let you know if it workse when I implement it.

Cheers, just had to get these thoughts down on "paper".

SN

[Chndru]

Nice thinking killroy. I think the major assumption is that vistors' preferences are fluid.

I can identify the long term popular ones, like the top 10, even though they might fluctuate up and down a little each day.

In general terms, the above statement holds true. So i dont know if dynamic prediction would be anything significantly different than your already established top 10, unless i entirely misread your post.

[killroy]

Let's put it this way:

I have 100 products. I have a top20, raltively stble, I have a top10, of which perhaps 5 are frequent customers, and I have a top 3, of which perhaps only 1 is always in the top three, or more accurantely, is in th etop 3 for more then a few days. The idea would be to have perhaps 1-5 run-of-site add links in addition to, for example, adSense, or even as adSense alternates, or completely instead of adSense if this proves more profitable.

These ads could run the long-term top5, with the risk of not catching the short term spikes. a simple quadratic prediction could say, ok, it increased by 40% yesterday, if this continues it'll be #1 tomorrow, so I place it in the ads, in preparation for the "rush" on that product.

Most popularity of widgets is actually more ralted to some ranking higher and some lower. But sometimes, a lower ranking one slips up into the top spot, without it's SE ranking cahnging. THAT's a short-term spike, and It'll be worth it the extra sales to catch those spikes.

SN

[Chndru]

to catch those spikes.

ha! now i see what you are trying to figure out. In this case, the time-window is extremely sensitive and temporal fluctuations (like time of the day, day of the month etc) needs to be taken in account to have any reasonably accurate prediction.