Advanced Statistics and Adwords

Mathematics of Adwords

[BradP]

I have seen someone write that an Adwords consultation is a good job for someone with a background in statistics. I have played around with adwords a little and I don't quite see how an academic experience would apply, other than giving you a gut feeling on some trends. It seems like like the adwords campaigns vary too much from day to day to apply any sort of rigorous statistical analysis. Can anyone comment on that?

[DiscoStu]

The greater the variation the bigger the sample needed is all. Stats will always vary day to day, but if you use enough conversions/clicks in your analysis you will get significant results, just a matter of getting enough data. How many conversions/clicks are you looking at and a across what length of time?

[BradP]

It was pretty long tail when I did it, so there were maybe like 2,000 clicks with 10 sales, or something like that, with a high variation on conversion rate over time. It was split between a few different products though. But I guess even if I had a nice big sample size, the only statistic analysis I can foresee doing is determining whether the mean commissions were greater than the mean advertising costs, and determining the confidence interval, based on sample variations, that they would remain that way.

[DiscoStu]

You can't run any kind of analysis on 10 sales - even arithmetic mean will have big variance. The problem isn't that adwords fluctuates, it's that your sample is too small. With enough data there are many ways to slice and dice data, you could for instance predict a 95% confidence interval for conversion rates for specific keywords and set individual bids accordingly. You can also analyze the effects of seasonality, time of day, day of week, country, language, match type etc etc. But these are more for enterprise level accounts. If you're looking at a smaller account the only analysis I'd worry about is if you're losing money or not (but 10 sales is still not enough to tell even that). If you have some knowledge in statistics you can however calculate at what point a specific keyword/match type is 95% or more likely to lose money in the long run. and then cut those.

But if it's a new account or new site, chances are that other factors will change (landing page, account structure, ctr) before any such test reaches statistical significance, which contaminates the data.

tl;dr - don't worry about any special analysis unless you're looking at 1000s of conversions

[RhinoFish]

a background in statistics should ground you in outlier analysis, minimum data set size, importance of causal contribution, segmentation, effective dissection and more. however, a consultant also needs a non-analytical side that gives her the freedom to float into new areas on hunches, not data crunching.

once sitting on a nest of baby keywords, apply testing and data - but to find them, don't be rigorous and mathematical. generally, explorers are a little foolish - if lewis and clark or columbus had analyzed their risks first...

lastly, statisticians often write analytical ads - which is frequently not the mode the ad viewers are in. if you're a number cruncher, read Cialdini so you can better use your analytical minds in an emotional world.
en.wikipedia.org/wiki/Robert_Cialdini

[LucidSW]

Having a background in statistics can certainly help in improving your campaigns. But PPC is first about marketing. If you know little or nothing about marketing, all the statistics will tell you is that you could be doing things better. It's like baseball where statistics help in evaluating players, what they did in the past. But there is more to that, such as a player's character (read Moneyball).

[Baylow]

This is exactly what makes finding a good PPC campaign manager so difficult. I'm looking to grow my team at the moment and having a complete headache finding someone to come in at entry level. You either get people who are too Quant and can't sell a product to save their lives, or you get people who are marketing/sales minded who couldn't wrap their heads around a regression analysis if it was spoon fed to them.

I've personally had a ton of luck with Economics majors, and investment bankers as a model for a good PPC manager. Though at this point i'm seriously considering splitting the roles of ad text creation and keyword/campaign management. It's just two totally different skill sets.

[arieng]

Robert Cialdini: +1

[BradP]

Great replies, thanks.

[Eurydice]

Baylow, you suggest to hire investment bankers to manage PPC? At what salary? And what bonus?

The original question was about statistics for GAW. I wait for the KW to get 1,000 impressions and then look at the CTR or conversion rate. With 1,024 impressions, you have a +/- 5% margin of error, which means the result (i.e., the CTR) is fairly reliable.

However, I agree with the general trend of the thread: it's more about marketing and sales than statistical analysis.

[Baylow]

No, not investment bankers.

College grads that have an interest in investment banking, and possibly those that have interned there.

[Renaud_Joly]

Things showed by sales figures are often end of live. Search is more about trends and alerts than probability.