Our bid management system allows our clients to keep full control of their campaigns while relying on robust algorithms and statistical analysis to manage bids.
Based on the objectives provided by our customers for example in terms of ROI or CPA, the system will use historical data to calculate the optimal bid for each keyword. Changes will only be made when sufficient data is available. The bids are changed directly in AdWords or Yahoo! Bing, where all changes are also logged. This provides full visibility on the bid management activity and also allows for roll-back.
To maximize the benefit of each keyword, our algorithms estimate the expected profit it can generate. The mathematical computations are based on historical data for the given keyword, and if insufficient data is available at that level, comparable data is used. Given the low volume of data generally available, the power of our algorithms is required to best determine the statistical relevance, determine the best set of data to use, and take calculated risks that will overall maximize the results.
The AdQuantic bid management system manages bids at the keyword level. Google AdWords and Yahoo! Bing allow advertisers to set different bids for each keyword, which is the most specific granularity level. Even two keywords that are very close from a semantical point of view can yield very different results, which makes keyword-level bid management superior. For example, we noticed that in French, keyword string “create an Internet site” had a much higher conversion rate than “create Internet site” for our client Overblog.
The following diagram illustrates this principle.
The diagram shows in red the percentage of costs and in blue the percentage of revenues generated by the 18 most important keywords within the campaign. On the left, which corresponds to the situation prior to optimizing, there are significant discrepancies between the investments for a given keyword and the subsequent revenues generated.
For example Keyword 1 generated about 11% of the revenues, whilst costing around 18% of the advertising budget before optimization. This situation corresponds to over-spending, making the keyword unprofitable. Note that it does not mean that the keyword is necessarily bad, but rather that bids were too high before optimization. Our algorithms therefore decreased the bid, reducing the sales generated, but most importantly taking it to a level where the keyword contributed positively.
Inversely, Keyword 3 generated about 18% of the revenues, whilst only costing around 11% of the total budget before optimization. For such a keyword, our algorithm will increase the bid, decreasing its margin percentage, but generating more volume, and overall profit.
Although this mechanism seems very simple, implementing it is complicated. First because the bidding mechanism means that we do not pay our actual bid for each click, but the one of the next highest bidder. Second because we pay per click, which means that to project revenues, we need to be able to estimate conversion rates. All of this is at the keyword level, where sometimes very little data is available.
Managing bids properly involves projecting revenue for each keyword. However, this revenue projection can be complex given the low volume of data available. A statistical analysis for each keyword is therefore required initially to determine the appropriate dataset to use: the one that applies best to our keyword at the moment we want to make the projection. Depending on the volume of data, our algorithm will, for example, determine the best time period to use, and potentially use data from other similar performing keywords. Also, this is where methods from quantum physics are applied because of the discrete nature of the data.
The analysis that our algorithms carry out forecast potential revenue for each keyword depending on its bid. However, this remains a statistical prediction: certain keywords will perform better than forecasted in a given time period, despite having the same potential as others, which did not perform as well. In the following period, it could be this second group of keywords that generate more revenues than anticipated. Our algorithms use mathematical models to take calculated risks on each keyword, which might or might not pay off at the keyword level, but generate performances at the campaign level. This mechanism allows the AdQuantic bid management solution to outperform competing technology, which restricts bids to well-performing keywords, and misses out on some revenue or conversions opportunities.
Performances of each keyword constantly evolve, due to external factors (seasonality, competitive activity) as well as internal factors (for example changes in pricing). This is why bids need to be regularly adjusted.
Except in some very specific cases, our algorithm never changes bids by more than 25% at a time. It verifies the positive impact of the previous change before applying a new one, hence ensuring that every modification contributes to better performances. This approach allows for a robust and secure bid management process, while generating significant improvements in less than 3 weeks after starting the optimization.
Our algorithms calculate the optimal bid modifiers to be able to adjust bids by taking into account seasonality factors (time of day, day of the week) as well as target devices (mobile devices for example).