Efficient Frontier Insights

Algorithmic Bidding in a Seasonal Marketplace

Optimal bidding for performance is a complex task for a Search Marketer. When placing bids everyday, the advertiser has to keep in mind the volatility in the marketplace due to changing search traffic as well as competition, matching algorithms that vary across search engines, and the changing needs of one’s business. For businesses in sectors such as travel and retail, seasonality is an additional factor that makes campaign management very daunting.

Seasonality refers to the cyclical patterns in the demand for various product offerings. For instance, winter clothing sales peak in the November-December timeframe after which they decline.

It would be very difficult for any human to sift through historical data on all keywords of a large campaign and identify statistically significant seasonal patterns. Further, in businesses such as retail, different keywords exhibit different seasonal patterns; summer clothes don’t sell in winter and vice-versa. Hence, to ensure optimal performance, seasonal keywords would have to be continuously identified and bid to appropriate positions.
The solution to this problem lies in building predictive keyword models that correctly estimate the expected revenue and spend patterns while taking seasonality into consideration. At Efficient Frontier, our predictive models detect such seasonal keywords within an automated framework. In addition to analyzing search traffic patterns, we determine seasonal keywords that have a high probability of improving the overall portfolio ROI and prioritize learning these keywords while keeping the advertising budget under check. This automated identification of revenue generating seasonal keywords is especially valuable to large and medium sized advertisers, since the typical size of their accounts involves creating predictive models on hundreds of thousands of keywords daily.

 
The example above shows the effects of incorporating seasonal models for a set of apparel retailers. The seasonality algorithm identified three types of seasonal keywords which were to be bid to higher positions; generic winter keywords such as sweaters, brands that make winter clothing, and finally brands that sold well during winter, perhaps due to the shopping season. The identified keywords were bid to appropriate higher positions. The results of the bid ups are shown below.

The Spend levels of all terms increased between two and five fold. At the same time, the ROI increased between 2 to 2.5 times. We note that once the seasonal decline is observed for these keywords, they will be bid down to appropriate levels so as to maximize the ROI of the advertiser.

Spend Trends in the United States

The analysis of spend trends in a sampling of our clients reveals the impact of the weak economy in the United States. We divided our client sample into (a) Large Spenders who spend more than $200,000 per month, (b) Medium spenders who spend between $50,000 and $200,000 per month and (c) Small spenders who spend less than $50,000 per month. Next, we compared the change in spend between Q3 and Q2 for 2008. The results are shown below.

The chart reveals that:

1. On average large spenders have increased spend by 5%, medium spenders have decreased spend by 4% and small spenders have decreased spend by -15%.
2. The standard deviation of the spend change was larger (by 20%) for the small spenders than it was for the large spenders. In other words, small spenders have responded more diversely on than the large spenders. While on average they have reduced spend, many have also increased their spending quarter over quarter.

The data shows that small spenders are being significantly affected by the current economic situation. Medium spenders have seen a slight decline in spend. The large spenders are bucking the spend trend perhaps with the aim to consolidate their current market positions.

Method of Analysis: We divided our client sample into three types based on spend levels. Next, we calculated the histogram of spend changes for all clients in each spend category. Empirical observations revealed that the distribution of spend changes was approximately a Normal distribution. With the calculations of mean and standard deviation, normal distributions were fitted to the histograms as seen in the chart.