* Update* With recent changes to Doubleclick’s data policies, user id has recently been removed from data transfer, which has crippled a number of analysis functions that have previously existed within the industry (not CUBED’s though!). This change has been explained very well by Jack Shearring on his LinkedIn post
As part of the development of CUBED, we never wanted to rely on our partners for the core interaction data. We therefore developed our own pixel system that integrates not just with Doubleclick, but also with other DSPs and display vendors. This has enabled us to maintain our tracking of impression led activity and consequently ensure that impressions are included as part of the consumer interaction journey.
The focus of this blog will be on display advertising (visual creative). However, a lot of the logic and calculations are also utilised in email impressions and direct mail impression logic within CUBED.
When a user is served an advert, a request for the CUBED pixel fires and we capture a number of data sets dependant on the DSP. In most cases the campaign, creative and placement are all captured, which allows us to understand which ad is being served to the user. At this point we also capture the device and IP led location, however, this is not utilized currently in the algorithm.
Each impression is assigned to an individual and our cookie (third party) is created, which means that all impressions will link up against a specific user id. If that person has already visited the brand’s website, or will visit in future, this cookie turns into a first-party cookie and all impressions against that user id are assigned to the visitor id.
In all circumstances, impressions and visits are then aligned into a single customer view. The single customer view is in the form of a time-stamped timeline in which we can see all interactions in the order that they happen. All data is anonymised and so, as much as we are aligning against a visitor id, we actually have no knowledge of who this user is. In some cases, for some clients, we are talking over One Hundred Million impressions a day, and it is therefore vital that the impact of that impression on a visitor/converter is taken into consideration.
Dealing with huge volumes of impressions requires a change in thinking when designing the algorithmic approach. We want to ensure that each impression in isolation has value. This needed to be scaled though, and this meant that we wanted to see impressions as clusters for individuals where multiple impressions of the same ad featured between visits. In CUBED we group common impressions with a count, which means that we do not have to deal with impressions on a unique basis while still proportioning value to the lowest granularity.
There are two core thought processes;
The simple way to showcase the value of a campaign is to show the variance in the conversion rate of those people who have visited the website and have seen a campaign, versus those people who have visited but have not seen the campaign. This gives an insight into the differences in buying behaviour in the people who have seen a campaign and those have not, thus providing an idea into the value of that campaign and its performance in driving conversions.
The algorithm performs this level of cluster recognition at scale comparing millions of rows of data where impressions either feature or do not. This trains the machine learning to understand which impressions have value and which are being served inefficiently.
Viewability has been a hot topic over the last few years and many platforms have their own viewability scoring to let advertisers know what percentage of their ads could be seen by potential customers. The problem with the lack of a consistent system is that if you run campaigns across multiple DSP’s or the logic of what is viewable is not fully accurate your logic for optimisations may be incorrect.
At CUBED we do not have any concept of viewability (on purpose). The reasoning behind this is that if a placement is not visible, it should not ever impact consumers. The data that we capture shows the change and propensity that each impression has. If a placement is consistently not having any incremental impact on the conversion journey, then it is likely that that advert is not fully or easily visible, maybe it requires extensive scrolling. This logic gets smarter over time but can easily highlight to advertisers those placements that are never driving any incrementality and therefore could save you some money! Equally, those great placements combined with great creative will be highlighted to drive incremental impression impact.
We are constantly improving and adjusting the impression impact within the machine learning to make it smarter. Working with both agency and technology partners enables us to take into consideration as many possible influencing factors at an impression level.