AD QUALITY

I have indicated earlier that the ranking used by both Google and Yahoo! is based not only on bids, but also on a measure of ad quality. In the simplest case, we can think of ad quality as the predicted click-through rate.

Think of a model where the actual click-through rate that an ad receives depends on both a position-specific effect (x_p) and an ad-specific effect (e_α). The simplest specifica­tion that the click-through rate for ad a in position p is given by e_a x

Given this multiplicative form, it is relatively easy to estimate the relevant values: simply put random ads in position p to estimate the position-specific effect. Once this is known, you can use the history of clicks on a given ad to estimate the ad-specific effect. One can also use various other predictors to supplement the historical data. In practice this is done using a kind of huge logistic regression utilizing nearly a trillion observations.

The ranking of ads is based on bids times ad-specific effects: b_a e_a. The bid is dollars per click and the ad-specific effect is clicks per impression. Hence b_a e_a is bids per impression: how much the advertiser is willing to pay for its ad to be shown to a user. The advertiser with the highest value for an impression is given the best position: the position most likely to receive a click. The advertiser with the second highest value per impression gets the next best position, and so on. Hence an ad with a high bid per click could be displaced by an ad with a lower bid if the high-bid ad had a low click-through rate. Assigning ads on the basis on b_a e_a maximizes the value of the impressions on the page, leading (potentially) to an increase in expected revenue.⁹

Just as it is important to determine which ads to show, it is equally important to deter­mine which ads not to show.

The decision of whether and where to show an ad should depend not just on current ad revenue, but on an estimate of how the ad’s relevance will affect future propensities to click. It is possible to model these choices analytically. Showing an ad today brings in a known amount of revenue but also has a probabilistic effect on future revenue by influ­encing the propensity to click in the future. Modeling these effects leads to a stochastic dynamic programming problem that offers a rationale for current practices and a guide to how they might be refined.

18.9