In the past month I have hear two very different approaches to cookie window, both delivered with equal amounts of hubris:
1. First is “we set our cookie windows to one day…and that is what we pay on”
2. Second is “we want a perpetual cookie.”
The first statement can lead to trouble because multiple touchpoints often contribute to an action, especially when richer-media and social media engagements are used.
The second statement can also lead to trouble unless advanced analytics are used to determine causality – eventually, any given user will consume a nearly infinite amount of impressions.
Below are the results on time-to-convert by publisher type, based on a study of Atlas advertiser data. One recommendation is to assign conversion attribution by publisher type, based on the expectations around time-to-convert. Search gets a day to land a conversion, real estate gets 27, etc. Of course, after building and running a model, an advertiser or agency should analyze the results to determine whether the net actions (or conversions) increase, or the marginal cost decreases. The analysis should determine “did the model work?” Next step: optimize the model.
With respect to cookie windows: with the right analytical approach, an infinitely long window can be used (cookie degradation notwithstanding). However, if you want to use an 80/20 rule: 35 days, which does not capture all the conversions, but does get to the median days of the longest conversion window.
Popularity: 3% [?]
Based on my experience, one really needs to build a probabilistic model based on actual ad history data to determine how to allocate credit to different kinds of ads. Heuristic allocation of credit based on what seems to work better is subject to substantial bias.
Implementing a multi-touch attribution approach will, over time and with regular optimization, increase digital productivity by upwards of 10% via better ad investment decisions.
One clear influence path that modeling corrects is search conversions. Frequently, search clicks and search conversions are heavily influenced by prior media exposure, but disproportionately take credit as the last ad.
Should also mention that rich media does not necessarily drive conversion any better than flash or static media.
Bryan – Thanks for contributing.
For my dear readers (and,um, me) could you illustrate the distinction between probalistic and heuristic modeling, perhaps by giving an example of each? Heck, if you can talk about some of your experiences with both (sounds like you have) I’ll post it as a “guest post.”
- h
Hi Harrison,
You bring up an interesting topic. There might be a couple of approaches we can take.
1) Marketing Mix Modeling: If you don’t have metrics about how a customer is exposed to various forms of advertising (both online and offline), we can do a macro level marketing mix modeling that looks at all media exposures (not at customer level) and its impact on sales or conversions. You can limit this to online as well.
2) Multivariate Analysis: Let us assume that we have all the information about customer level exposure to all forms of online media through a tool like Atlas. Now, this allows us to build regression analysis to identify key drivers to conversions. There are also other models such as stochastic grandient boosting and decision trees that can give more insight into how different media exposures are driving customers conversions.
H…
I would absolutely love to write a guest post. I’ll reach out to you directly.
But just for starters –
Atlas “last ad” model is an example of a simple heuristic model (in this case giving one ad 100% of the credit) ….”most recent click within the click conversion window gets 100% credit. If no clicks, then the most recent ad viewed in the view conversion window gets 100% credit”
Another heuristic model might be a basic straight-line sharing model like “I will share equal credit across every ad that the consumer saw or clicked within x days before conversion.” You could even choose to weight, say, clicks as 10x and rich media as 2x for this heuristic model, and thus grant more credit to clicks and rich media because you believe that these are more likely to have had a positive effect on conversion.
These would be a heuristic (rules-based) models – often chosen for simplicity or because they align with manager’s expectations for how consumers are influenced by media.
Probabilistic models would be a special case of analytical models (such as those that Naga describes). Models based on analyzing media influence and deriving quantitative approaches to sharing credit (versus rules of thumb).
Multivariate regression can be used to derive the relative probability that certain ads or ad combinations will result in conversion. These probabilies can be used to apply differing amounts of credit to each ad in a consumer’s exposure history.
In my experience, this sort of analyis can improve efficiency (conversions per media dollar) by as much as 10% or more (assuming flat overall ad spend before vs after). This improvement comes from better recognizing the influence of earlier ad exposures in setting up the likelihood of a consumer converting on a client’s product.
I’d be interested in different ways readers of your blog have approached this problem – and also whether they see more accurate attribution as a way to improve advertising performance, or whether it seems too daunting to consider.
Someone named Emily had a good comment on multivariate testing, and it got caught in my spam filter and I accidently deleted it. Emily – if you are reading this, please repost!
- digitalCMO