You Ask, I Answer: Attribution Modeling for Social Impressions?

In today's episode, Maggie asks, is there a way to build a custom data-driven attribution model to include social impressions such as Facebook and Instagram data? Yes, there's always a way. Both require machine learning, and both are very possible, and something that actually I've done very, very recently for a couple of Trust Insights customers. Let's look first at the traffic-based one. Because that one is, I think, probably the best overall blend of social and Google Analytics data.

Um, Google Analytics gives you path information in their multi-channel funnels uh API, which is how people interact with your company uh on outside sources before they come to your website. Um for example, if they view a Facebook post and then they click on it, and then they view an Instagram post, they click on that, and they view an email and click on that, you would see in the API in the data, you know, Facebook, Instagram, um, email, right? As the chain of evidence that leads to a conversion or doesn't. Now, there's a machine learning technique called Markov chain modeling that you can use to pull apart those paths and re-sequence them. Um version, which ones don't, and then what the differences are.

So in the that same example, Facebook, Instagram email. Now let's say that leads to a conversion, and then you have Instagram email as a different one, uh a no conversion. In that very, very simple example, you can see by taking Facebook out, the conversion fell apart. And so if you do this, well, you have a machine do this thousands or millions of times with all the conversion data in your analytics instance, you can then apply weighting to say, guess what? Facebook Facebook matters because you know, in in 87% of conversions that didn't have Facebook, they fell apart.

Right? So that's the traffic-based one. And what I do is I take that data and then I condense it down to something a little more understandable. Let's actually bring this up here. So this is an example of a digital customer journey analysis, which is what a Markov chain model is.

Um by source and medium, from top to bottom and from left to right is the order in which people do things. Because again, same back to that earlier example Facebook, Instagram, email. Facebook's at the beginning of the journey, Instagram in the middle, email at the end for that conversion. So if you just essentially tally up those conversion pathways, you get this chart where you can see here uh impact BND is is uh 8.6 of our conversions, uh percent of our conversions are begin there, same with LinkedIn, same with um Facebook, then right at the end of the awareness journey is Google Organic Search. That's a big chunk, 27%.

It's actually relatively healthy because anything over like 50 and you're you're in danger of being too vulnerable on one channel. You see all the um engagement channels, and then at the last half of the last third of the journey, you see email, social, uh, Twitter, things like that. So, this is one way of sequencing that data in order to get at uh how different social channels play into attribution. Now, this is if you have everything wired up and you're passing as much data into Google Analytics as possible. If you just have the impression data itself, you're doing a type of modeling that uh essentially is called uh view-through modeling.

And Facebook has like, for example, view-through conversions, and what you're trying to do is show a mathematical relationship between the number of impressions that your social content has made and any sort of impact in the rest of your analytics. So, what you need to do in this example, and in if you're just trying to work on the impression data itself, is you need to build a very high-resolution table of essentially every interaction, ideally broken down like the hour level instead of the day level of impressions, clicks, whatever other metrics you can get out of Facebook, whatever metrics you can get out of Instagram, you put them all in one giant table, and then each each network's um dimensions are their own columns of Facebook impression, Instagram impressions, Twitter impressions by hour of day, by day of we uh by day, by week by month. You should have, you know, for example, uh, one day should be what 1440 rows per network. And then you have to you call the Google Analytics API and bring in things like branded organic searches, organic searches, direct traffic where there is a lack of attribution, um, your total number of goal completions and things like that, and you will use a technique probably like uh gradient boosting machines, um, extreme gradient boosting is the preferred method used a lot of the time for for analysis like this to say here are the factors, here are the dimensions that drive the outcome we care about, like conversions or branded organic searches, whatever KPI you're trying to benchmark against, and then the software will go through and mix and match every possible combination. The thing that's important is that you need a high resolution table.

If you were to do this at like month level, you would get useless data because the sample is just too small. You really need as granular data as possible. Google Analytics can spit out I believe you can go down to the second level in Google Analytics data. And for Facebook, Instagram, I'm I think you can get hour minute um out of uh out of those APIs. Again, you put it all together, run your gradient boosting machine, extract an answer.

Um, and then in this instance, here's the catch. Because you're dealing with essentially uh at the end of the day, it was very uh very fairly fancy correlations. Um you need to then build a testing plan to test the correlations. If Facebook matters the most, right, and maybe it's Facebook videos. If you have like factors to the the different types of content, you definitely want that in there.

Facebook videos matter the most. Then you build a testing plan and you run it and you do lots of Facebook videos and see if as the number of Facebook videos goes up, do you see a commensurate increase in the target metrics that you built with the gradient boosting machine, like branded organic searches. You can do more Facebook videos, absent anything else, do you see a a a corresponding proportional increase in branded organic searches? And you test that, you test that, you make sure it's not statistical noise. You test for statistical significance and then at the end of the testing period you say yes, we can prove because we've done the the testing properly that Facebook videos increases your whatever KPI you care about.

So those are the two primary ways to build this kind of data driven attribution model. Um ideally as much of your social data should come in as click traffic as possible, but I know that's not always possible. Use gradient boosting machines using machine learning to uh assemble that and shameless plug if you want help doing that, because done it a lot, uh feel free to reach out. Happy to to have that conversation. As always, please uh leave your comments in the comments box below.

Please uh subscribe to the YouTube channel and to the newsletter I'll talk to you soon. Want help solving your company's data analytics and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.