Find solution
What would be a reasonable price for a CPM ad?

Does a second-price auction for ad slots penalize improvements in CTR prediction algorithms?

  • Say a company (Google for instance) has a pretty decent CTR prediction algorithm that puts ad A on #1 because it has a predicted CTR of 1.5% with a $1 bid, and ad B on #2 because it has a predicted CTR of 1% with a $1.20 bid. If we simplify "Quality Score" to be equal to CTR, ad A has "ad rank" (bid x quality score) of 1.5, and ad B has ad rank 1.2. As Google usually is said to charge "ad rank of the person below you / you quality score" in this case the charge would be $0.80 (1.2 / 1.5). However, if in reality ad A is actually much better than ad B, suppose in the extreme case that ad A in this situation has a 100% CTR and ad B 0%, the company would still earn $0.80 if it uses its 'bad' CTR prediction, but it will earn nothing if it would have an optimal estimate of the actual CTRs. In other words, for a second-price auction, it seems that the required CTR prediction algorithm should be just good enough to ensure a decent ranking of the ads, but absolutely no better, as that will just lower the charges of actual clicks that will happen. Or am I missing something here?

  • Answer:

    Let's start with an idealized second-price auction. In that auction, the winner is charged the CPC (cost per click, or "bid") of the first runner-up. So in this case, if ad A is clicked, Google will earn ad B's bid of $1. In the idealized example, Google earns the same money either way. This is instructive because it illustrates a general point: Google's click prediction doesn't affect the actual CTR on the ads on the page. The predictions exist so that Google can predict which ads will get clicked. They don't affect what actually gets clicked. In general, if Google makes bad click predictions, what happens is it shows worse ads in a worse order, resulting in a worse user experience (a terrible outcome) and less money earned (an unfortunate side effect). In your example, the differences in prediction don't change anything in the real world -- they don't change which ads are shown, in which order, what the advertisers bid, or what the users actually clicked on. So the only effect is that Google was more blind to what will actually happen. Now, as you correctly point out, in Google's auction, the quality of the ad (for which CTR or "ad rank" can be a proxy) impacts its price. Which means, yes, it is possible that improved click prediction could result in lower short-term earnings. Here's an example based on your question: If ad B had predicted CTR of 0%, it would not get shown at all. This means ad A will get charged the minimum bid on that query -- let's say it's $0.05. This is indeed less short-term money for Google. But let's look at what Google has accomplished here: They hid ad B from the user, which greatly improves the user experience, because ad B is a terrible ad. They improve the long-term value of their ad real estate by not cluttering it with junk. They can expect higher revenue in the long term as a result. Advertiser A is a happy customer because she is not having her prices jacked up because of some crappy competing ad nobody will click. If ad A has a 100% predicted CTR, then it is simply a wonderful search result, paid or otherwise. Google is lucky to be able to show it to users. God forbid advertiser A runs out of budget because Google got greedy, and Google is forced to remove that result from their page, substantially worsening the user experience. So when all is said and done, reducing immediate earnings to improve user experience is a pretty great tradeoff for Google to make. And to their enduring credit, they make it without blinking pretty much every time.

Harry Glaser at Quora Visit the source

Was this solution helpful to you?

Other answers

As the question notes and Harry Glaser's answer explain, it is possible that improvements in CTR prediction can reduce revenue and it's possible to construct extreme examples (such as the one you describe) in which perfect CTR estimation results in minimal revenue. You can think about changes in CTR estimation having two primary effects: 1) They change the order that the ads are shown in. 2) They change the ratio of the second highest ad's eCPM (expected cost per thousand impressions) to the first highest ad's eCPM. When making CTR estimation more accurate, the first effect should generally be positive for revenue but the second effect can be either positive or negative. But the following is not a valid conclusion: In other words, for a second-price auction, it seems that the required CTR prediction algorithm should be just good enough to ensure a decent ranking of the ads, but absolutely no better, as that will just lower the charges of actual clicks that will happen. This is because it's just as possible for CTR improvements that don't change the final ranking of ads to increase revenue. For example, suppose there are two ads  bidding $1 per click with true CTRs of 4% and 3%. If the previous CTR estimation algorithm was correctly estimating a 4% CTR for the first ad but incorrectly estimating a CTR of 1% for the second ad then correcting this error will increase revenue even though it won't change the ranking of the ads.

John Hegeman

I agree with Harry. To elaborate on what he said, there are a couple of options a company has that are superior to intentionally getting bad CTR estimates. Introduce a squashing parameter like Yahoo used to: http://www.theregister.co.uk/2010/09/16/yahoo_does_squashing/. By changing the parameter you can provide more "price support" to sparse auctions, or by moving it the other direction you could make the price lower / position better for high CTR ads. Change the reserve prices. By changing the minimum price to be displayed, you can avoid the situation letting an ads price drop down to zero when you remove a bad CTR competitor. Show more ads somewhere else to make up for the ads you didn't show on this impression. Google/Bing have a trade-off between user happiness and revenue. Too many ads makes people unhappy, but generally increases revenue. If you remove low quality ads from some searches, and add some better ads to other searches, you might be able to increase user happiness and revenue at the same time.

Rob Donnelly

Here are a couple of points to consider.First, ad position is associated with CTR - the theoretical assumption (and most empirical evidence I'm aware of) is that all else equal, the ad with a better rank gets a higher CTR.Therefore, Google would need to "standardize" ads' historical CTRs in regards to their positions to eliminate this effect and determine "true" ability of an ad to generate a click-through.Not sure if they're doing this, so I always recommend new advertisers to over-bid to reach higher positions and therefore get into the self-reinforcing cycle of high CTR/position.Anyway, this endogeneity makes it tricky to predict CTR without accounting for positions.Second, Google has an incentive to get the highest rank ad clicked because that has the highest second-price as per its true bid. Therefore, the question of prediction is important. Using historical data, as they do now,  seems to be the best way to go about it. Like you stated, zero % CTR is not probable and even if that would happen, Google would quickly learn the ad is not performing and drop it altogether.So, even if this sub-optimal pricing would take place in some instances, in the long run the system adapts to performance difference between advertisers to maximize its own revenue.

Joni Salminen

CTR influences Quality Score and not the other way around. The second price auction model, although not without its shortcomings is still the best  method to deal with large numbers of bidders, of which many are new to the auction process and thus have no quality score history.  Plus it is simple to explain. And 100% CTR ? Biggest CTR value I've ever heard about (over a decent search volume) was around 35%.

Louis Havriliuc

Related Q & A:

Just Added Q & A:

Find solution

For every problem there is a solution! Proved by Solucija.

  • Got an issue and looking for advice?

  • Ask Solucija to search every corner of the Web for help.

  • Get workable solutions and helpful tips in a moment.

Just ask Solucija about an issue you face and immediately get a list of ready solutions, answers and tips from other Internet users. We always provide the most suitable and complete answer to your question at the top, along with a few good alternatives below.