Some of our advertisers are interested in a more detailed summary of how our predictive models operate, allowing them to more finely tune our platform to meet their needs. We have provided an outline below of the "standard" process for generating bid recommendations for an advertiser:

The key assumptions that we use to make predictions include the following:

• The advertisers has a maximum CPA target of C (this can be changed through our Goal settings)
• The advertiser wants to maximize profitable volume on each keyword. This means that no individual keyword should run at an expected CPA higher than C.
• Conversion rates are binomially distributed. When the keywords have gathered a minimum amount of data (configurable through the Bidding Preferences page), we use the average conversion rate.
• We use the Clopper-Pearson method to calculate the confidence interval based on the per keyword data for keywords that don't have enough data to meet the minimum click threshold you have specified.
• Conversion rates are independent of position. (This has been verified by Google, and confirmed through a number of AdBuyer analyses)

Using the above, we calculate the Expected Revenue for each keyword:

• Exp. Rev. = P[Conv|Keyword] * CPA

We will recommend setting the maximum bid to the Expected Revenue, though we won't recommend any increases to bids already in position 1.

A few extra notes:

• For users with variable revenue sources like e-commerce providers, or multiple CPA-driven revenue sources, we use a similar approach but include the average transaction value on a per-keyword basis to calculate the expected revenue per click. 
• We only use one-end of the confidence interval to change bids. When we increase bids, we use the lower-bound and use the upper-bound to decrease bids. In plain English, this means that we tend to "trust" your initial bid settings until we're proven wrong. If you want to quickly improve ROI, we recommend using the "Long-tail" option that will make changes faster.