Jul 262012

Beyond Average: An Analysis of Search Ad Position

An excerpt from my piece over on Search Engine Land today, if you missed it there:

Among the scores of data points search marketers have at their disposal, perhaps no other metric is as shrouded by mythology and misconception as ad position.  Despite careful studies showing, including one by RKG in 2006, and confirmation by Google in 2009 that conversion performance doesn't vary much by position, you will still hear pronouncements, even among enterprise level practitioners, that some specific position on the page generates exceptional performance.

This piece is not meant to rehash that argument, per se.  Instead, we'll examine how attributes of the current ad auction including: a Quality Score calculated at auction, broad matching, dayparting, increasingly personalized results, local competition and a changing competitive landscape, all conspire to scatter an ad across the listings regardless of anyone's best efforts to "own" a particular position.

The Ambiguity of Average Position

Up until about a year ago, when Google introduced a ValueTrack parameter to pass along position for each ad click, this analysis wasn't easy to do at scale.  Those using Google Analytics have had a limited view of click position for some time, but for others, short of scraping the SERP, which may or may not have been above board, the best view of ad position was previously Google's impression-based Average Position, usually viewed at the daily level.  While the name itself clearly highlights it is just an average, and thus our ad may appear across multiple positions, it offers no insight on the extent to which that is occurring.

For example, if an ad has an Average Position of 3.5, did it appear 50% of the time in position 3 and 50% of the time in position 4?  Or, did it appear in positions 2 through 5 25% of the time each.  Furthermore, Google's Average Position tells us nothing about which positions actually drove traffic to our site.  For that we'll need to assess the click position data we have via the ValueTrack parameter.

Actual Click Position Versus Daily Average (Impression) Position

Our sample here consists of over 120K ad clicks that took place over the course of a single day on about 50K unique keyword instances, i.e., identical keyword phrases in different ad groups or with otherwise different settings are viewed as separate from one another.

To keep the chart below readable, we've rounded the daily average impression position to the nearest integer.   We've also grouped anything over position 10 in one bucket for both the click position and the average position.  This includes any clicks from ads appearing on page 2 or higher.

The size of each circle here represents the percentage of all clicks for that daily average position that occurred at the given click position.  Or, the same data in table form:

So, for ads with an average position of 1, 95% of clicks happened when the ad actually appeared in position 1, 3.6% happened when the ad appeared in position 2 and a small percentage of clicks occurred in lower positions.

A number of interesting elements jump out here, including:

  • First and foremost, it's clear that ad clicks really are taking place across a wide array of positions regardless of the average daily impression position.  The spread escalates the lower the average position is, but is apparent throughout.
  • Except for in position 1, most clicks do not take place at the same position as the average impression position.
  • Click position is shifted upward from the average impression position.  This makes intuitive sense as click-through rates increase as we move up the listings.
  • We see a relative peak in clicks at position 3 and a relative drop off at position 9.  The former likely has to do with ads moving from the side listings to the top of the page at that point.  The latter may be a function of appearing below the fold.
  • It isn't until we get to an average position of 4, that the top position doesn't generate a plurality of clicks.

For the full analysis, including a breakdown of Search Network and match type effects, check out the original post.


3 Responses to "Beyond Average: An Analysis of Search Ad Position"
Dylan says:
Very interesting data here. Thanks for putting it together Mark. If you have two other data could potentially make this more actionable maybe? 1) CPC for different positions for a given average impression position. 2) Percentage of impression share for different positions for a given average impression position Armed with those two extra data, we could calculate whether more clicks resulted from higher "average impression position" result in incremental revenue/conversions that are profitable. So we will be really looking at the "portfolio" performance of a "average impression position" by analyzing the ROI for each actual ad position. For example, if the "average impression position" is 3.5, we can calculate the Cost, Return and Margin for the actual ad position, 1, 2, 3, 4..and 10, then add up the Margins for each position to arrive at the combined Margin for a give "average impression position". With these tangible data in front of us, we then can make more accurate decisions on : do I want more revenue or more margin, or a sweet spot in the middle, and where that sweet spot is going to be. But still, there are the dynamic side of ad auction and market changes and etc... But I am curious, is this something RKG factors into its proprietary bidding system? Loyal Reader, Dylan
Dylan, To my knowledge we still have no way to see our CPC for individual clicks, so we can't calculate how CPC might vary for the same keyword appearing in different positions at a given average impression position. If we could, it would certainly open up a number of interesting possibilities, but it's not necessary for predicting marginal returns on bid changes. RKG has long argued that ad position is basically irrelevant to our bid choices. Instead, what's important is to be able to accurately predict the value of traffic at a given bid and how traffic levels and quality (i.e. margin per click, sales per click, conversion rate, etc.) will change as we change our bid. Past performance and data from tools like Google's Bid Simulator allow us to make these predictions, which are largely indifferent to position because of some of the factors I explain above. Specifically, that a bid increase is not bumping us from one position to the next, but only changing the distribution of clicks across positions and thus moving us along a continuum of traffic levels. If our predictions for marginal CPC and return are accurate, we can decide whether to shoot for more revenue or more profit. And yes, RKG factors all of this into our bidding technology.
Dylan says:
Hi Mark, Thanks for the reply. I've read your related posts on "position bidding" and I am on the same page with RKG. Therefore, your reply makes total sense. What my earlier post was assuming by adding up the margins for "clicks across positions" will not equal to the margin calculated using the total cost and total value for a given average impression position. My initial instinct was that since different click positions will have different CPC and impression share, its margin weight will influence the resulting margin for a given "average impression position". However -in reality- the result is the same because the margin weight for each click position is already factored in when calculating its own margin. i.e. Margin(click position 1) + Margin(click position 2) + ... = Margin (given average impression) When I was writing the post, I got too caught up with my visualization and forgot using simple math :) It reminds me of how interestingly math works sometimes and why conceptualizing math formulas is an important part of math knowledge. Dylan

Leave A Comment