The goal of paid search management is to predict the future. To do that, PPC tools need to look at the past. The question is, How far back should they look? Let’s first review what the systems are looking at. In most bid management systems, the formula for calculating the actual (not maximum) cost per click (CPC) is:
CPC = (AOV × CVR) ÷ ROAS
Average Order Value (AOV) and Conversion Rate (CVR) are based on sampling historic performance, while Return on Ad Spend (ROAS) is typically a fixed value specified by the retailer. Most technologies will allow you to specify different ROAS or A/S goals for different keyword segments. Over time, AOV and CVR change—and in some cases very quickly—due to many factors. One critical factor is seasonality. Take the example of DVDs. During most of the year, demand for DVDs is relatively flat, so AOV and CVR don’t move much. But DVDs are a popular Christmas gift, so demand for DVDs—and thus CVRs—begin to climb in mid-November, hitting a peak just before Christmas. Then CVRs plummet.
Most bid management systems only look at the recent past to make bid optimization decisions and are blind to upcoming trends.
Where Bid Management Systems Go Wrong
Without knowing that this is a recurring pattern, how do bid management systems compute the “correct” CPC for DVDs in January, or any other month? The short answer: they don’t. The majority of systems look at a fixed length of trailing data—a look-back period (LBP)—which typically ranges from 30 to 180 days. Returning to the DVD example, if it’s January and the LBP is 60 days, most systems will predict that the progressively climbing CVRs observed during November and December will continue to stay strong, and thus in response keep CPCs high. That would obviously be the wrong thing to do. Likewise, if the system looked at six months of data prior to November, it wouldn’t predict the upcoming seasonal sales spike and would thus keep CPCs too low during the holiday shopping season. Again, the wrong thing to do. What, then, is the right LBP?
Determining the Appropriate LBP
The LBP needs to be just long enough to obtain statistically significant data—no longer, no less. The longer you look back, the less recency the data has, which reduces its potential to accurately predict the future. But to obtain sufficient data, the LBP might need to be quite long. The key issue is that one size does not fit all. The appropriate LBP will vary based on product type and by keyword. It should be based on how much recent historic data is available, normalized by what seasonal patterns are expected for this type of keyword. The LBP for DVDs, for instance, may be to the same period one year prior. For other products with high-volume keywords, it may be much shorter. For example, if October historical information (for this year) indicates that a DVD is going to have a conversion rate of four percent, and we know that this type of DVD shows a doubling of conversion rates from October to the end of November, we can predict that this DVD will have a conversion rate close to eight percent.
The Long Tail, Latency and Roll-Ups
Adlucent’s Deep Search™ software platform factors in several other issues that affect LBPs. Consider long-tail products and their keywords. Clicks on keywords for products far down the tail may be so infrequent—say, once every 30 days—that it takes months to compile sufficient data for predictive purposes. In fact, the majority of keywords don’t generate a lot data within a short period of time. Deep Search thus determines the appropriate LBP on a per-keyword basis. Other systems struggle with such low-trafficked keywords because certain short LBPs would indicate the keyword hasn’t sold any volume at all. Another factor is a latency effect due to AOVs. Consumers research (shop for) higher AOV products for longer periods of time. For these products, you need a longer LBP, otherwise it may seem that keywords aren’t effective when in fact they are—if you look at a sufficient period of time. While “past performance in no guarantee of future results,” it’s still an important variable in predicting what will happen tomorrow. The trick is to appreciate that how far you look back can have significant impacts on the success of paid search programs. Determining unique LBPs on a keyword-level basis and factoring-in issues such as seasonality and AOV latency will also dramatically improve the model.