I'm excited to share this next interview with you, because I always enjoy chatting with my friend, Mike Ryan. If you haven't already, make sure to catch the first full length episode here: Episode 2 - Attribution Concerns & Pitfalls
One of the things I love about this conversation with Mike, is how clearly he thinks through ramifications of technology, as a pro-tech person working at a software company. Give this episode a list, and ponder not simply the attribution models that can be used in your Google Ads account, but the ramifications of utilizing them incorrectly... and the ramifications of Google making decisions in an increasingly black box world.
In this episode, you'll hear the rest of the conversation Mike and I had that we couldn't fit into our full Core episode on attribution as we all ponder digital attribution together.
Article Mike References in the post: Want to improve your measurement? Get a grip on incrementality
Mike Ryan is Portfolio Strategist at Smarter Ecommerce (smec). Mike combines years of retail operations experience with expertise in digital marketing to deliver insights with a bias toward business outcomes rather than campaign outcomes. Mike has additional experience in data visualization, automation technology, and leading innovation projects.
Mike Ryan (00:03):
You want to have something that's realistic enough andthat's beneficial to you and that's not going to mislead your business. Also,you're not killing yourself doing it.
Kirk Williams (00:14):
Don't use your attribution model as the gospel truth. Useit as a directional tool to help you make decisions.
Mike Ryan (00:23):
People want personalization, but they also want privacy.It's tough to achieve both of those things simultaneously.
Chris Reeves (00:36):
Welcome to the ZATOWorks PPC Ponderings Podcast, where wediscuss the philosophy of PPC and ponder everything related to digitalmarketing. Today's show is a bonus episode of our full interview with theportfolio strategist at Smarter Ecommerce, Mike Ryan. Mike was heavily featuredin our second official PPC Ponderings episode on attribution. If you haven'tlistened to that episode yet, go give it a listen. Otherwise, please enjoy ourbehind the scenes conversation with Mike.
Kirk Williams (01:07):
All right. Thank you so much, Mike, for joining us. I knowyou and I have chatted a bit about a lot of things. Probably mostly shoppingads. We've chatted with Patrick Gilbert who will be a upcoming guest on afuture episode. Yeah. So thanks so much for joining us here, excited to getchatting about attribution.
Mike Ryan (01:28):
Yeah. Yeah. Likewise, thanks so much for having me andyeah. I mean, it's been a while since we chatted face to face, virtually faceto face, so that's awesome.
Kirk Williams (01:38):
Yeah. Yeah. Cool. So let me hop in. I think, I've givenyou a little bit of context about what we're talking about, especially ourprimary show as we kind of piece that together, but we're talking attribution.So we're going to talk both kind of the philosophy of attribution and thenhopefully get into some practical things. Like what are the models? How shouldwe think about the models? Do you have a favorite model? Things like that. Sohow would you define attribution? What is attribution?
Mike Ryan (02:07):
Great question. I mean, attribution in end effect, youbasically are talking about finding a series of events that all contributedtogether toward a specific outcome. For example, a purchase in an online storeand then those events could be ad clicks or different kind of views. You justwant to assign value to those because this is how you can help measure theeffectiveness of your advertising and which pieces mattered and how do they allwork together or how do they harmonize.
Kirk Williams (02:46):
Okay. So why does that matter? Why should we care aboutattribution?
Mike Ryan (02:51):
It's super important. I mean, I think it's one thing that'sreally attractive about digital marketing in particular is that there's it'sunleashed sort of a world of data here to work with where if you think abouttraditional media in the past, it was a bit more haphazard, the way you couldconnect that back. There are ways of doing that too. With digital marketing,it's just a great way to help connect the value of marketing back to thebusiness and also to help understand what's working, what's not working, andwhere you can improve. It ends up being a really important tool to understandwhat's going on in emarketing.
Kirk Williams (03:35):
Yeah, it definitely has an important role because withoutit, well, maybe without it, it has been where PPC has been a lot over the pastfew years, which is maybe by default more of like a last click model, right? Ithink attribution has probably evolved, but even our understanding of it hasevolved as part of that.
Mike Ryan (03:58):
Totally, I mean, you said it right there with the worddefault for last click, because that's the way that it's been up until recentlyand the power of default settings. So automatically a certain amount of themarket or of advertisers out there are operating on a certain worldview becauseattribution models, there's a couple, that word right there model is important.It's not a one-to-one representation of reality. It's kind of a perspective onwhat's happening. It's a way of understanding and you'll get a differentperspective with last click than you will with other models. This has been thedefault view that people had for a long time is last click.
Kirk Williams (04:42):
So Google Ads as the time of this recording, oh, I don'tremember, it's been maybe a few weeks, has announced, hey, they're movingdefault from last click to DDA, data driven attribution.
Mike Ryan (04:54):
Kirk Williams (04:55):
What is last click attribution? Why don't you start therejust in case they are people listening who are just not familiar withattribution models at all, which is totally fair. So what is last clickattribution and then maybe give your thoughts on should we be concerned aboutlast click attribution? Are there concerns with it? If so, why? Yeah, go aheadand hit us with last click attribution knowledge.
Mike Ryan (05:19):
Yeah, sure. I mean maybe if it's fine, I just zoom outeven a touch more. So when we talk about attribution, there's basically thesesimplistic models and fractional models. So last click is a type of simplisticmodel. What that means is that you'll basically have one touchpoint that'sgetting assigned all of that value that we talked about before. You could getthe first click or the last click. Then there are these fractional models whereyou can have partial credit getting assigned to different touchpoints so thatyou can distribute across more touchpoints.
Mike Ryan (05:58):
Then, well, we can get into DDA after that. But last clickis basically what it sounds like and where the last ad that was clicked getsassigned the value for the conversion. So that's a good approach in some ways.It's like a safe approach kind of when you first think about it, because ofcourse there's a pretty high correlation, a clear relationship between thatlast click and the conversion.
Mike Ryan (06:31):
It's just easy to connect the dots there. So it's pretty safeto say that there's a relationship there and this is what a model is all about.It's about trying to find a relationship between different occurrences orhappenings, events. Yeah, on the other hand, it's quite simplistic. The riskthere, the classic risk is that these last touchpoints are typically going tobe lower in the funnel. So that could be in terms of the way your campaigns areset up and the type of maybe creative that you have out there, or the type ofkeywords that you're bidding on or however.
Mike Ryan (07:12):
So when you are assigning the value all the way there,it's going to put a bias in your view against the upper funnel. You might startto be missing the value in the upper funnel and not realize the value the upperfunnel is actually driving for you. So it can strangle these really important brandingmeasures over time or awareness measures that were actually instrumental. Youdidn't even know about it if you're just looking at last click.
Kirk Williams (07:40):
Nice. Perfect. Okay. Let's go back to some philosophy typestuff and then we'll get to some models in that, I'd love to talk more aboutDDA and that.
Mike Ryan (07:50):
Kirk Williams (07:50):
What are some ways the attribution is misunderstood maybeor misused? Where can it go wrong?
Mike Ryan (07:59):
Yeah. As I mentioned with that word, that word model isjust so important because there's this famous quote from George Box that allmodels are wrong and some are just more useful than others. Super important tokeep in mind with attribution, it's very applicable here because we'll neverknow the mind of the individual consumer and we'll get back to this with DDAtoo, I guess later, but you don't want to overfit to that either. It's justabout finding the relationship. It's just about finding an approximation ofreality that is useful to you.
Mike Ryan (08:42):
So a misunderstanding of attribution is maybe that this isactually reflecting objective reality, that can't be the case entirely.Attribution, it's always going to be just one view of reality and thesedifferent models can have value in terms of offering you differentperspectives. Attribution often gets confused with incrementality. This I thinkis something that comes up a lot and they're kind of related topics.Incrementality is more the question would a given outcome have occurred withoutmarketing pressure or advertising pressure, for example. Attribution is moreabout understanding the value of these touchpoints and how the picture all cametogether.
Mike Ryan (09:36):
So you'll have people confusing these terms and using themkind of interchangeably. There's a great blog post, by the way, if you haveseen this one from Avinash Kaushik from Google, definitely worth a look by theway. Don't have to include that in the show.
Kirk Williams (09:52):
Do you have more info on it so I can make sure I get thatexact post because would include that in the show notes.
Mike Ryan (09:57):
Oh yeah, definitely. So it's a Think With Google article.I can send you the URL later.
Kirk Williams (10:02):
That'd be great.
Mike Ryan (10:03):
Yeah. But these are solving for different problems, incrementalityand attribution. Yet I think that they get confused together sometimes. Yeah. Ithink those are two ways that attribution gets misunderstood.
Kirk Williams (10:20):
Then what are some challenges to attribution? What makesattribution more difficult than maybe someone would expect?
Mike Ryan (10:31):
There's no more challenges at this point becauseattribution in digital marketing will typically have dependency on cookies, forexample. This is becoming increasingly a topic in terms of privacy and theconcerns that get raised there and the regulation and there's this tensionwhere people want personalization, but they also want privacy. It's tough toachieve both of those things simultaneously, but that's kind of targeting is adifferent topic a bit.
Mike Ryan (11:13):
So, but attribution gets challenged definitely by thecurrent regulatory landscape right now. Attribution gets challenged by theamount of data that might be required or the modeling, especially if you wantto do custom attribution or get into data driven attribution. There's thisdifference between if we go back to that quote from George Box, he talks abouta couple of ideas. He says that all models are wrong. Some are more useful thanothers. Another thing that he discusses is that you basically don't want to,how should I say this? I should look up his exact quote.
Mike Ryan (11:57):
You don't want to kind of over-engineer a model becausebasically the effort that gets involved outweighs the benefit of that effort.So this is a thing why a model like, for example, time decay or position basecan be useful because they offer you a approximation of reality. At the sametime, you don't have to put that much effort in them. They're kind ofcomputationally. This is a heuristic approach, which is basically, it'seducated guessing. You want to get a reasonable approximation of what's goingon without making it rocket science. Bees do this even. Bees have to solveincredibly complex problems, like finding out what's the most efficient route topollinate as many flowers as possible.
Mike Ryan (12:50):
This is even challenging for our computers. It's calledthe postman's dilemma. So they're not going to find the most efficient route.They're going to find a reasonably efficient route that was computationallylightweight. Yeah. So there's this whole topic here with attribution that youwant to have something that's realistic enough and that's beneficial to you andthat's not going to mislead your business, and also you're not killing yourselfdoing it. That's an advantage of Google making tools available, for example,because they have more computational power and they have more data. So they canmake these pretty sophisticated approaches available to anyone.
Kirk Williams (13:32):
That was great. You even kind of answered my nextquestion, which was about what are some limitations to attribution?
Mike Ryan (13:38):
Kirk Williams (13:38):
Yeah. It's funny. I like your little quote in there.People want privacy and they want great targeting at the same time and thosebasically work against each other. Right? It is going to be interesting to seewhat happens as privacy just continues to kick in. So I know you're over in theEU.
Mike Ryan (13:57):
Kirk Williams (13:57):
We in the US, we're always behind the EU in terms of lawsand that sort of thing. Even I think [inaudible 00:14:02] connected with thisbriefly on LinkedIn and that about even the difference of how they see things.EU, there's an interest in preserving some sort of competitor relationship,that sort of thing. With the US, it's just like, hey, has a specific law thathas already been written, been broken? If not, continue. Right?
Mike Ryan (14:25):
Kirk Williams (14:26):
Then everyone's like, crap. Wow. That blew up in ourfaces. I guess we're going to have to write more laws. Right? So kind of adifference in the way that EU and US look at things like that, which so that'swhy some ways you have privacy always lagging behind, but still moving in thatdirection. Yeah, as we talk about things like probabilistic models ofattribution, that's what they're doing. So yes, they have a mountain of dataand yet they're still at some point using that mountain of data to make betterguesses, especially as more and more of that data starts disappearing.
Kirk Williams (15:05):
It doesn't mean that we just completely throw this out thewindow and let's go back to guessing. It doesn't mean that it's completelyuseless. That is why we should be looking into this stuff though. That's why weshould be probably thinking through, are there ways we could use multiplesources and multiple attribution models just to think through this stuff. Thenat the end of the day, man, it is so important that people, if anything, what Iwant to do is to convince people, don't use your attribution model as the gospeltruth. Use it as a directional tool to help you make decisions.
Kirk Williams (15:43):
In some ways, in my opinion, it's okay for a really greatmarketer to make a decision that an attribution model maybe doesn't fullysupport, but because they have just insights into this is the audience we'retargeting. There might be more issues with getting data on them for privacy orwhatever than others. Here's ad creative that we just know has worked in thepast. Attribution tool isn't telling us this is going to be a home run, but youknow what, we're going to go ahead and set aside this experimental budget andtake a risk with that. I think that's okay. So anyways.
Mike Ryan (16:22):
No, I agree, totally. I think it's tricky because we allwant to work with data and we all want to be data driven and there's a certainamount of almost social pressure to go that way. It's easy to derive experienceor creative intuition has gut feeling. I just think that there does need to bea balance there because we forget that machine learning we're trying to, yeah,okay, it can do things that humans cannot do, but we are pretty smart too.We're way smarter than machines in a general way. Machines can be very good atspecific tasks.
Mike Ryan (17:06):
Yeah, I think that idea that attribution is not the gospeltruth, that's so right. The whole thing, when we talk about limitations ofattribution, attribution only knows what it knows. There are things that arehard to measure. There's data that is out of scope. Sometimes there can be avalue by the way in anecdotal evidence too. If you see something happening onceit's probably happening more often, which is something that is maybe getting tobe a controversial statement in 2021, in 2022. I don't know. Yeah, if we're atthe philosophical level, there it is. So, yeah.
Mike Ryan (17:48):
Another thing that about challenges to attribution that wedidn't mention is the topic of bias. This is connected to the limitations onthe one hand, because if you're talking about Google attribution, okay. Ifyou're using one of the 360 products, I think they have Marketing Mix Modelingin there, which is a next step above data driven attribution. Gets into what'soccurring across your channels. There can be a tendency for Google to biastheir own touchpoints because they know more about them. That can happen onaccident.
Mike Ryan (18:23):
There can be a tendency for Google or anyone, Facebook,whomever, to bias their own touchpoints because it's financially attractive forthem to do that because it helps prove their value to advertisers. You mightfind that you have conversions that are getting claimed by multiple channelsand is where it becomes really challenging to take a bigger picture across yourchannels and harmonize that and try to under understand how they're allsupporting each other, how they're team players.
Mike Ryan (18:53):
There are also human biases that can come in place. Therecan be an incentive for marketers to want to prove their value. There can be aconfirmation bias in there that they want to prove that something they think isgoing to work is working. If we take these gut feelings and transform them intohypotheses and take a structured approach, then you can test yourself. So that'sa whole other approach to it.
Kirk Williams (19:22):
I love that. Let's just be cautious. Let's not throw thebaby out with the bath water and get rid of attribution models or completelysay there's no benefit in them. That's even a factor platform bias, marketerbias. Anyone can make any data say what they want to and just kind of presentin the right way. Attribution models certainly are no different, especiallywhen it's the platforms themselves. So Facebook, I remember a couple years backand this finally changed, I think. People's understanding evolved, but Iremember I would almost template my response to our clients who would have somesort of, whoa, Facebook is just completely dominating and outperforming Googleat the time.
Kirk Williams (20:08):
Basically I'd have to kind of introduce them to the ideaof the different attribution models being used, especially since, again, atthat time Facebook was using view throughs, Google wasn't. So you would havethese seven day click-through windows, one day view through windows or it mighteven be longer than that for Facebook. So Facebook just looked like it wasdoing just these unbelievable things directly from, in their platform. We can'tcompare that to Google because it's literally different ways that they'relooking at that data. So it's just not an apples to apples comparison. That'sokay. Let's be aware of that.
Mike Ryan (20:46):
Yeah. You said it. It's apples to oranges and this is achallenging thing of, when you're dealing with different channels, they mightjust be providing different kinds of metrics. Then how do you equate those canbe pretty challenging.
Kirk Williams (21:03):
Okay. So what's your preferred model within the Google Adsplatform? You get a client. What model do you like to set their conversions at?
Mike Ryan (21:16):
Well, at this point, for me, the answer is easy. It's datadriven attribution. I think it's a capability that people have wanted for along time and, or, yeah, I don't know to what extent Google has generated thedemand for this, but it's a wonderful capability to have though in my opinion.It would not necessarily be within the reach of in-house teams or agency teams,because it is a challenging topic. You almost need a third party specialist orplatform specialist to solve that technology problem.
Mike Ryan (21:55):
Yeah. About data driven attribution, we talked aboutfractional attribution models before and data driven attribution is one ofthose. So there are different approaches here, because there's a lot of waysthat you can build up a data driven. It's not like it's all one thing. It'skind of a category, but you basically want to model the value the differenttouchpoints had based on the data that is available, rather than based on anassumption going in like other kinds of rule based ones. Yeah. Those are humanrules. You say, okay, we're going to evaluate attribution in this fixed waybased on kind of human logic or human rules. The data driven attribution isjust going to let the data talk and say where the value should go.
Kirk Williams (22:55):
Okay. So if an advertiser doesn't have DDA yet accessibleto them, do you have a secondary preferred model that you would choose?
Mike Ryan (23:04):
Well, let me start by saying that kind of my leastfavorite just about is last click or first click. I just think that these aretoo simplified. What I like instead of those would be a position basedattribution model, which is going to favor the first click and the last click.I think that these are two really important touchpoints in any customer journeybecause you can say that that first click is somehow the known impetus or theknown first touchpoint that has a very important role. Whereas that last clickis what kind of convinced them to make a purchase or caught them when they wereready.
Mike Ryan (23:47):
So these are two really important touchpoints. So I likeposition based for giving attention to both of them, because it shouldn'tstarve your upper funnel in the same way as last click. Time decay is kind ofinteresting too. It's like an extension of last click, but it just gives someweight to previous touchpoints and the further back the touchpoint goes theless weight that it gets. These just in my opinion, they're nice rules to havein place, but they pale in comparison to what a data driven attribution modelwill do.
Kirk Williams (24:21):
Do you set different models for different conversion?
Mike Ryan (24:25):
Yeah, that's a great question. By the way I come mostlyfrom a ecommerce background. I think it can be totally valid to set differentmodels for different kind of conversions. At the end of the day, DDA should bea way of solving, arguably any use case. There are different challenges ordifferent kind of use cases that can pop up. I mean an interesting thing thatjumps to mind that I've seen in a B2B environment, I don't know if I've everheard or seen somebody write about this before, but we call it a jackpot click.Tell me if you've seen this before or heard about this.
Mike Ryan (25:09):
Basically within a cookie window, what happens is somebodyclicks an ad, then let's say, yeah, it's a B2B user. Let's say they're workingfor an appliance repair place. Okay. So they're ordering appliance parts on aroutine basis. Maybe they order parts three times weekly or something likethat. Or maybe they build up a cart over the course of a week and buy it.
Mike Ryan (25:38):
Then let's say that either they bookmarked a link withGoogle click ID in there, or they start typing in the domain name, let's sayacme.co or acme.com or something. Then their Chrome browser fills out to thepage that they were on. So it should be direct traffic, but it still is gettingthat click ID in there. So you get a click that just has an absurd amount ofconversion value that it racks up over 30 days or whatever the cookie windowis. So it's a jackpot click and that's something that needs human attention.
Kirk Williams (26:17):
That's super interesting. I mean, I've been aware of thepossibility of exactly the Google autofill or bookmarking that specific clickwith the GCLID attached. But I don't know if I've seen that beforespecifically. I'll have to dig in a little and see if I can find evidence ofthat in our account. Yeah, exactly. It'd probably be more though, as you said,when there is a lot of return on purchases is where you would really notice it.
Mike Ryan (26:52):
Kirk Williams (26:52):
You'd have to have an account like that.
Mike Ryan (26:54):
Yeah. I don't think it would tend to occur in B2C. I'dhave to think through if there's a realistic scenario for that, but we have B2Bclients where we've seen this occurring. Yeah.
Kirk Williams (27:07):
Gotcha. Interesting. That is interesting. Yeah. Probablymost of our clients are B2C. Interesting. Okay. So let's see. Let's see. Let melook through my questions here.
Mike Ryan (27:18):
Are there a long sales cycle too?
Kirk Williams (27:23):
Yeah. Why don't you talk about what about with long salescycles? You had just talked through how you've seen some evidence on B2B of oneinstance. Are there other things that someone with a longer sales cycle,specific challenges, things that they should probably think through?
Mike Ryan (27:40):
Yeah. The thing that happens with a long sales cycle, Imean, everybody will see this to a certain extent in their campaign traffic,the delayed conversions or conversion lag. There's going to be a certainpercentage of users who will click on an ad and buy right away. Then there'sgoing to be a certain amount of people who maybe spend some time making adecision, building a cart, comparing options on other websites. They might completea transaction hours later, days later.
Mike Ryan (28:13):
What gets tricky is when you're in, let's say for example,a high value industry with these larger conversions or something where peoplereally need to consider awhile before they make a decision, is that they canlapse outside of the cookie window. A delayed conversion in the short term,your performance is going to typically look worse at first and improve over thecourse of a month while those conversions come through and get attributed.
Mike Ryan (28:48):
It's a different topic when your cookie window lapses andit appears like a new click. It appears like the old click didn't occur. So,you can choose what your default cookie window is up to 90 days, and we'll seehow this technology's going to develop in the future with a flock in all ofthese topics in a post cookie environment. But it's definitely a challenge orsomething to be aware of when you have these kind of longer buying cycles.
Kirk Williams (29:17):
How would changing the attribution model impact smartbidding? Does it? And if so, how?
Mike Ryan (29:24):
So I don't know exactly what goes on in Google'stechnology, if they have some ways of resolving this. I would expect definitelythat it would have some kind of an impact. So the thing is, if you've beenmeasuring in a certain way for the last year, it depends what's going on behindthe scenes with Google. But typically let's say you want to predict what a bidshould look like. You're basically going to try to predict the conversion rate.
Mike Ryan (29:57):
You're going to try to predict the average order volume.You'll want to do this on the basis of historical data. The thing is when theconversion that you get assigned is changing due to attribution and you make abig cut like that, the lookback window that you're looking at has beenmeasuring one way. Then it's a question is, can you make valid bids andpredictions on this new model, this new way of looking at things?
Mike Ryan (30:29):
So what I would do, if I would have the power, so fulldisclosure. We have a bidding tool. We do have the power to set the lookbackwindow or exclude dates or ranges, which would be super cool to have this inGoogle too. Because if you know there was something anomalous going on in youraccount or you made some major changes that you could give Google a heads upthat they shouldn't look at that period.
Mike Ryan (30:53):
I am under the impression that Google, I think they'reaware of a long term picture, but I think that they're optimizing on, they havesuch a data volume and their algorithms are so efficient that they are lookingmuch more at a short lookback window, that they're looking on a shorter timerange. I think that they can adjust to changes like this pretty quickly. Iwouldn't be surprised if it triggers a learning phase for any algorithm,including Google's because it's a notable change in the account.
Kirk Williams (31:26):
Yeah, totally agree. For the sake of getting a littlesnippet on that, can you define what a lookback window is?
Mike Ryan (31:33):
Sure. A lookback window is basically a period of time andthe data that's inside of that time series that you can use to make decisions.So for example, you might look at a 365 day lookback and to understand the bigpicture, what is my conversion rate over time in a rather stable way? You canget a very stable average with that.
Mike Ryan (32:03):
If you're looking at the last seven days, for example,it's going to fluctuate a lot more and there are advantages to both things. Ifyou look at last 365 days, you can get a broad sense of seasonality, forexample. If you look back at last seven days, you can get a picture on trendsand things that are merging, what's happening now currently. So typically youwant to look at both things. Does that make sense?
Kirk Williams (32:32):
Yep. That's perfect. Is there anything else that youwanted to make sure that we hit on? Was there a specific question I didn't askthat you kind of had a great answer prepared for? What's a final thing weshould think about on attribution?
Mike Ryan (32:48):
Do you want to get into the different types of DDA outthere? I can't get into super technical detail, but anything like that?
Kirk Williams (32:59):
Absolutely. Yeah. Why don't you go ahead and fill us in alittle bit more on data driven attribution.
Mike Ryan (33:06):
Yeah. So basically with data driven attribution, these areevidence based models and they're going to assign fractional value to differenttouchpoints based on the evidence that supports that. DDA is not one singlemonolithic thing. There are many different ways of approaching the topic. Asimpler way of approaching it would be like a regression model. Thedisadvantage of that is that it wouldn't be the most advanced model compared toothers.
Mike Ryan (33:42):
You can have these game theory approaches, if you've heardabout Shapley values in attribution, for example, this is referring to gametheory. The way this works, when we're talking about game theory, this isbasically a way of understanding rational decision making and the interactionbetween decision makers. So we'll talk about cooperative game theory there.Just imagine that you have people who are working to win a game together andyou want to give value to each of those individuals for their contribution.They have a same goal, which is to win this game. They have to cooperate. Sothis is one approach to attribution.
Mike Ryan (34:23):
The game here is winning a conversion and the individualsare channels or different touchpoints. These people also can form what arecalled coalitions, but you might think of it as just teaming up together. Sothat's display plus email plus paid search, or just paid search plus display,or email plus paid search, or different combinations. So the Shapley value isjust a way of quantifying this and modeling this.
Mike Ryan (34:53):
There's also these Markov chains, which you'll hear aboutsometimes. It's just a statistical model where each step is only dependent onthe last one. So there are these random variables in there, but you just kindof build each step on the last one. Because these things tend to look atcorrelation, you can get into causality as well, like cause and effect througha counterfactual model. What would have happened? Stuff like this. The importantthing about DDA is that you're trying to create a balanced model of what'shappening.
Mike Ryan (35:33):
If we talk about something like last click, this isdefinitely sort of under fitted. It's too generic. It's not that realistic. Ifyou'd have a over fitted model, this would be basically trying to get everythinginto there and including errors and variation end up getting into the mix. It'snot actually as reliable. You need to find a balanced model where you have arealistic and generalizable view on the patterns that are occurring or therelationships that exist in between your marketing channels, in between your adtypes and how this is impacting your customers.
Kirk Williams (36:13):
So in something like DDA, is it utilizing all of the abovefor solutions to figure out how it's going to compile all that? So Shapley,Markov counterfactual. Or is it choosing one of those? Is the DDA just utilizedthe Shapley model?
Mike Ryan (36:34):
It's a great question where I don't know if Google hasreally in what extent they've disclosed how their DDA works. I'd have to checkthat out. I don't know off the top of my head. Now I want to look.
Kirk Williams (36:49):
I wouldn't be surprised if they didn't because I know thatthey hate to be gamed. They like to keep things close to their chest that theydon't need to give away. But yeah. Interesting.
Mike Ryan (36:59):
Good question though. I don't know exactly what's going onunder the hood at Google with that.
Kirk Williams (37:05):
Why not? One of the most private corporations in theworld. Cool. Well, hey, thank you so much. I don't want to take up any more ofyour time.
Mike Ryan (37:17):
Absolutely, Kirk. Thanks for having me. I'm a happylistener of the podcast. So it's really an honor to be on here. Thank you.
Chris Reeves (37:27):
This has been a bonus episode of the PPC Ponderings Podcast.Keep checking back for more interviews and our next full episode. If you likewhat you hear, please consider sharing this with your network, leaving us areview on Apple podcasts. Until next time, may the auctions be ever in yourfavor.