Fortune Tellers and Paid Bid Management
Keyword bid management for a paid search campaign can be accomplished in different ways. Are you trusting a "fortune teller" to guide your campaigns?
Keyword bid management for a paid search campaign can be accomplished in different ways. Are you trusting a "fortune teller" to guide your campaigns?
Keyword bid management for a paid search campaign can be accomplished in different ways. Campaign managers should be wary of those methods that require them to become a “Fortune Teller,” or at least recognize the situation for what it is.
At its simplest, bid management can be done manually – Fortune Teller 1.0 – to identify patterns and “guesstimate” outcomes from bid manipulations. In this case, managers often rely exclusively on historical data as inputs into their decision making, adjusting manually for trends, cycles, bumps and dips.
Sometimes, this technique uses the engine interfaces, while at other times Excel is the ancillary weapon of choice to combine data across engines, chart the data out, stick in some “adjustment calculations”, and extrapolate (roughly speaking) the outcome of a bid adjustment.
In some cases, the Fortune Teller (especially versions 2.0+) can be remarkably successful using historic data and manual extrapolations to do keyword bid adjustments. In these cases, the campaign manager is indeed prophetic. But as the campaign size grows (number of keyword-match types) and the “reaction” of keywords to bid adjustments does not conform to simple extrapolations, the Fortune Teller begins to get over-loaded and performs sub-optimally.
In fact, it is possible that even Fortune Teller 2.0+ can be overwhelmed and make BAD forecasts (and decisions) that result in deteriorating campaign performance. In these cases, it is natural to exploit the bandwidth and sophistication of computerized technologies.
In its least sophisticated form, computer processing power can churn through many keywords and implement rules created by the campaign manager. “If CPA > $45, then decrease bid by 10%.” You can well imagine the Fortune Teller doing this keyword-by-keyword, albeit in a torturous way using the engine interfaces. So, instead the engine data is downloaded, and computerized rules are put into place. The campaign manager explicitly creates rules in a computer interface that says “If CPA > $45, then decrease bid by 10%.”
These rules can, of course, process many thousands, if not millions, of keywords in very little time. Furthermore, if the rules are cleverly designed to run in tandem or sequence (rule layers), then it’s possible to do some pretty complicated bid management. (Certainly, more than can be done via the engine interfaces.) At the very minimum, a paid search campaign of any size or complexity should be using rules based on historical data. But oftentimes, this minimum standard of computer use does not perform well.
In general, the use of historical data to extrapolate into the future is often difficult. First, the extrapolation must account for many factors that influence paid search metrics, such as trends, cycles and seasonality, “shocks” (such as landing page failures, pricing changes), or other important external events (e.g., media mix). Second, it is even difficult to decide WHAT historical data to look at. The last week? Month? The last 6 Mondays? Are those data points predictive of what happens tomorrow (Monday)? Finally, does the rule actually predict (the best) bid move? Or is it directional (up/down) at best?
Thus, the use of rules in combination with historical data is like driving a car by looking in the rear-view mirror. It can be an unnerving experience, even for Fortune Tellers.
The use of rules running off historical data should often be replaced by rules (or even optimization algorithms) that run off of predictive model forecasts. That is, in order to account for the many factors that impact paid search metrics, it is necessary to build predictive models that explicitly (via statistical equations) incorporate the above mentioned inputs (trends, seasonality, price changes, media mix, etc.).
Given those inputs, the predictive models can forecast the influence of a bid level. Even more importantly, the predictive models can incorporate DIFFERENT BIDS, and therefore produce bid scenarios. Now things get interesting, because the campaign manager can choose amongst a suite of keyword-bid combinations. To do so, they can use rules, of course (running on the forecasted scenarios).
Or even better yet – they can implement a bid optimization algorithm that goes through all the keyword-bid combinations and finds the “best ones” given budget constraints. Can the Fortune Teller (v2.0+ at least) do that? I’m afraid to say, “no.”
A properly programmed computer with accurate forecasts and optimized bid selections will outperform the campaign manager using historical data. This is particularly true if the campaign contains at least several thousand active keywords and/or complex interactions between bids and key search metrics. (Okay, it might not be a bad idea to have computer and human “checks” to make sure the bid selection process is not running amok = “semi-automated bid management.”)
In summary, beware of Fortune Tellers. Most campaigns should be running predictive models in order to obtain bid impact scenarios. Additionally, those scenarios should be processed systematically in order to obtain the “optimal set” of keyword-bids.
Running rules on forecasted data is at best a meat-axe approach to bid management, even when guided by a skilled Fortune Teller. Anyone responsible for the performance of even a semi-large or complex paid search campaign should know that Fortune Tellers (even when armed with tables and graphs) will often lead you astray.
Pat Stroh is VP of analysis & decision support at SEM agency Impaqt. He spearheads analytical initiatives and decision support activities for Impaqt’s search marketing campaigns, with a special focus on ROI-optimization.
We report the top search marketing news daily at the Search Engine Watch Blog. You’ll find more news from around the Web below.