How do you calculate the organic lift for a specific cohort?
The total projected ROAS for a cohort is the sum of the projected ROAS from attributed installs and the projected ROAS from the estimated additional organic revenue.
As an example, let’s say we want to calculate the total pROAS for an iOS campaign on Facebook targeting US users on 6/1/2020. We have the following data:
|Spend||Projected revenue from attributed installs||Projected revenue from estimated organic installs (organic lift)|
The projected revenue of $1500 comes from summing our pLTVs of each user attributed to that campaign on 6/1/2020. The projected organic lift of $500 comes from our Organic Lift model.
The “pure” paid ROAS for this cohort is $1500 / $1000 = 150%. The total pROAS, including organic lift, is ($1500 + $500) / $1000 = 200%.
The same logic can be applied to any cohort.
How do you solve for signal-to-noise problems for clients with a lower overall amount of marketing activity, or volatile data?
It is not possible to know exactly what drove an organic install, and so we rely on statistical techniques to estimate OL from noisy data.
|Data Challenge||Impact||Our Approach|
|Low marketing spend||Low paid signal||Rely on OL from similar clients advertising on similar networks. If a client is not spending much on paid advertising, there is little point in calculating a small lift on total revenue|
|Large traditional marketing spend burst (i.e., TV ads)||Organic installs fluctuate sporadically masking the effect of digital spend||Traditional advertising will likely affect organic installs in multiple categories (i.e., iOS and Android) in a way that is uncorrelated to increase/decrease in marketing spend toward those categories. Then, our algorithm would reject that fluctuation as a systemic signal, not one that can be attributed to a specific network.|
|Poor attribution||Paid installs are incorrectly labeled as organic||From a strict modeling point of view, paid installs coming in as organic actually improve the signal to noise ratio and make it easier to estimate OL. However OL then is inflated since some paid installs come in as Organic. In a real data ecosystem where misattribution is common, we feel that OL should account both true organic AND misattributed paid installs, as the latter are still the result of paid UA activity.|
|Low campaign volume||Small cohorts don’t provide statistical power to estimate OL||As described above, campaign level Organic Lift revenue is the product of the statistically most likely number of organic installs AND the organic revenue those organic users are expected to bring. These two quantities can be estimated at different granularities to assume statistical significance. |
Our experiments have shown that significant difference in OL are found at the network level and probably reflect the ad serving dynamics of, say, Facebook’s wall versus Google’s YouTube. Thus we typically calculate OL at the network-platform level and then apply that learning to smaller cohorts.