You Ask, I Answer: Attribution Modeling for Facebook Campaigns?

In today's episode, Anna asks for attribution modeling. What model do you use on your Facebook campaigns? So Facebook offers uh substantially similar attribution models as Google Analytics. If you're familiar with the models in Google Analytics, first touch, last touch, linear, time decay, position-based, um, data-driven. Most of these models are have a Facebook analog.

Facebook's attribution models are unsurprisingly all centered around uh the social network and the advertising that you do on it. Uh it they are okay. Um they tend not to take into account uh all the other things that their campaigns could be doing, and they tend not to take into account um interactions between campaigns outside of Facebook. Um so just be aware that there are some pretty substantial limitations, uh especially if you don't have the Facebook uh analytics uh pixel set up and and properly working on all of your other digital properties. Umsurprisingly, Facebook's attribution models tend to paint paint Facebook in the best light possible.

Uh if you look at how uh things like the data-driven attribution model works, it only works for Facebook's properties. So Facebook, Instagram, uh the and the campaigns that you run on those uh platforms. And so uh again, you don't really get a very good picture of what is happening. In terms of all those models, which one to use if you have to use Facebook's attribution modeling, uh, is gonna depend on the uh the type of customer journey you have and how m how along the path is to conversion. When you look at models like first touch and last touch, um, if you have a very transactional relationship with your customers, they come in, they do their thing, and they're gone, right?

And they don't hang around, they don't engage, uh, they are not loyal to you, um, you are a commodity to them, then you know, first and last touch work fine. You know, if 99% of your transactions uh have a customer journey length of like one, meaning they come from one source, they do their thing and then they leave, then last touch is fine. You also don't actually need attribution in those cases because you know that you know 99% of your transactions coming from one thing, they come in and they do their thing and they leave. Um if you have a uh a more complex sales cycle, um, and one in which uh there's a lot of decision making, maybe multiple stakeholders, maybe uh a lot of consideration, then out of the box, time decay is the least bad model. Uh, as long as you set the window long enough to accommodate you know almost all the transactions that are possible.

So if it takes 60 days for someone to make a purchase, your your uh window uh of attribution should be at least 60 days, ideally 120, right, to cover all the outliers. That said, again, Facebook is not going to give you a comprehensive look at how it functions with the rest of your stack. Um for that, you're better off using Google Analytics attribution models and looking at Facebook's performance in that. You'll want to make sure that you are separating out uh paid social media traffic from unpaid uh in Google Analytics, get that set up in your analytics settings first so that uh you don't mix organic and and paid traffic together together. That's a a critical flaw that a lot of folks uh overlook and then use again the Google attribution models um choosing one of those again same idea you know first and last touch are are great if uh if you are are are transactional uh linear models where it's just even credit distributed across all channels is appropriate in environments where uh you have absolutely no idea um what's working and what's not and it's a good way to start experimenting and to see like you know is there a substantial difference in uh credit given in a linear model compared to say a a time decay model if there is then you may be starting to suss out that some channels are more important than others the best models to use um are ones that are not built into these pieces of software um there's one uh Google offers it in in their attribution product um but the useful part is really only an attribution 360 which is the the paid version of the product uh they use what's called Shapley game theory it is a machine learning model that uh looks at how models how channels change based on when a new channel enters the mix, and the analogy that is often used for that is a poker game.

Uh you have four players sitting at a table, they are all betting certain amounts. When a fifth player sits down the table, um, not only do you look at how the fifth player bets, but you also look to see at the behavior of the other four players to see how if how their behavior changes. You know, if some guy who shows up is clearly the sucker, uh, and a whole bunch of you know the other channels all start, uh the other players will start uh betting more heavily. Um Shapley Game 3 is applied to your digital marketing channels. Um if you add a new channel in, do all the other channels' performance change in a positive or negative way?

The model that I have found works the best, and again, this is something that requires uh a lot of uh technical experience uh to do is called Markov chain modeling, where um it behaves like uh a basketball game. Uh you know, if if uh Steph Curry is on the on the court, um you know he's gonna you know shoot the three-pointer, and he would be the example of sort of the last touch. But the question is who assists him to get the ball to him, right? Um in those cases when you're looking at uh who's valuable on the team, you need to figure out how who assists him the most, who assists the assist, etc. And you can build a model of um who is most effective and where the same thing is is done in Markov chain modeling in uh in machine learning for marketing channels.

Is there a channel that assists more? Um, even if it isn't what gets the last touch in. Now, again, less helpful if 99% of your conversions are one touch and done. Um, but if you have steps, if your product is something that requires consideration, um, then it's a really good model to use. So, time decay, good for uh it's it's the least bad of the models that are that are in the box.

Um Facebook's data-driven attribution model, nobody knows what algorithm they use, so I would be very reticent to recommend that. Uh first and last touch if your product is purely transactional and there's no customer loyalty whatsoever, nobody even ever comes back to your website after the after visiting once. Um and use the machine learning models if you have the technical and statistical capabilities to do so because that will give you the best answers possible based on what's happening. When you can, use Google Analytics instead of Facebook Analytics and uh and see how it goes. If you have follow up questions, leave them in the comments box below.

Subscribe to the YouTube channel and the newsletter. We'll talk to you soon. Take care. 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.