Blog Post

Using multiple regression for pay-per-click bid management


Chad sent me an interesting blog post today where the analysts were discussing multiple regression analysis as a tool for improving pay per click (ppc) bid optimization. This is intriguing to me as we are always looking for ways to improve our client's paid adverting (ppc) campaigns in the search engines.

But after giving the post a read, I realized that it was pretty much a simple primer on linear regression (multiple linear regression), explained in a search marketing context. It did not really give anything away that was at all helpful... But it got me thinking about this. Here are some high level thoughts on the idea of multiple regression for ppc bid optimization in a paid search advertising campaign:

1) Yes - With the right data, you could certainly run a linear regression to develop a model of the relationship between bids and clicks (or conversions). The data requirements would be serious, but not impossible.

2) To me, the big question is "so what?" When I have a regression line that helps me predict clicks from bids, will that really save a ton of work? I don't manage ppc campaigns directly, so I plan on consulting with our ppc analyst team to see if they disagree (and have a hunch they do).

3) The bigger issue is that using linear regression to define a relationship is a very different thing from finding maximum points of optimization in a paid search / ppc campaign. Very different. My feeling is that our current analyst-driven experiment approach is actually finding "soft spots" (or points of maximization) within the bid / click / conversion relationship. Of course, our process (which relies on the analysts intuition and ability to see complex mathematical relationships is inherently inefficient.... But on the flip side, if you flatten these relationships out into straight lines (which is what multiple linear regression does), you would miss these points of optimization.

4) I have to assume that the secret sauce actually lies in some other aggregation function, or the way you would test various scenarios. Of course, this is they type of statistical kryptonite that one is not likely to expose in a blog post. :-)

If anyone has ever tried this, I would be very interested in how you define "optimization." I could see that if you had developed, say, 200 linear regression models (corresponding to 200 keywords), you could then iterate through many different scenarios to see which one yields the most... There *might* be something there... (and I think there's a name for that, but I forget because it's been a while). I guess it's time to dust off my stat books and call a few grad school stat geeks!