So What? Applied Data Science in Marketing

Well, howdy howdy everyone. Happy Thursday. Welcome to So What the Marketing Analytics and Insights Live Show. I am joined by Chris and John. Today we are talking about applied data science in marketing.

And this is a topic that people have a lot of questions about. And don't you worry, Chris, I have got a bundle of common questions about data science for you, especially as it applies to marketing. So those are the things we're going to cover is how data science applies to marketing, um, how to look at various reports and interpret them, and the most important functions in marketing data science. So Chris, where would you like to start? Gosh, I don't know.

I think um one of the first places to start is that we need to contextualize data science. Uh and actually probably should explain it first. Data science is the application of scientific principles and techniques to data. So it is just science applied to data. Just like you have chemistry as a science, biology is a science, things like that.

Right. It's not some arcane art. There's no sacrificing of of things other than maybe you know, uh a a mouse that you break or whatever when you're when you're computing stuff. And there's nothing mysterious about it. It is a lot of math and some coding and stuff.

And that's what I think most people find somewhat off-putting about it, is that it is a fairly technical thing to do. Um even in the beginning, like if you're not comfortable, say like Microsoft Excel, it's it it you're gonna have a bad time. Um so it is one of those functions that uh you need to have some aptitude for, even if you don't necessarily get into the techniques, but it is good to know what it is and what it does. Again, the part that I think is most important for everyone to understand is it's not magic, right? There's no there's nothing secret uh about a lot of common data science techniques.

What makes it different is that very often these techniques which are developed in fields like finance and biology and medicine are not common in marketing, and so as a result, it does seem like these these crazy arcane arts. So I think that's uh the probably uh the first place to start. The second is like I said, putting it in context. Data science is at its core, it's a practice, right? Which means that it belongs to the same structural framework as everything else, um, which would be the 5P framework, right?

So you have your uh the purpose, why are you doing the thing? The people, who's doing the thing, the process, what is the thing that you're doing, the platform, uh, how are you, you know what tools using to do the thing, and then the the performance. Uh, did you do the thing? Well, data science is just a set of techniques that enables each of these stages. So I have a couple of common beginner questions for you.

Uh so number one, um, you know, we talk about things like marketing analytics or a marketing analyst. Is a marketing analyst a data scientist? And then the second part of that question is if I've only ever used Excel, if I've never coded something before, am I a data scientist? We can answer that both questions at the same time. You are doing data science if you're fundamentally performing the function of science, which is to uh test hypotheses, right?

To prove or disprove something. So if you are pulling in a whole bunch of data and there's no hypothesis, you're just cranking out a report. Here's what happened last month. You're doing analysis, you are an analyst. If, on the other hand, you have a hypothesis, say like uh sending out more tweets on Tuesday results in higher conversions.

And then you go and test that hypothesis by any mathematically valid means, you are now doing science, and that would make that data science. So it's not the tools, it's not the technology, it is the process that you're using. So again, we go back to our our our 5P platform. Data science is all about the process, uh, whether or not you're doing it. Okay.

Um, I think you know, that's an interesting way to look at it, um, because I think that there's a lot of confusion or even just misunderstanding about what data science really means, and that it's not an approachable thing that you can bring into your marketing practice or your business in general, because you need to hire specifically a data scientist, uh, which is you know probably a good move uh because it's someone who's going to be holding the rest of the team accountable to doing these practices. However, if you want to practice data science, then you need to make sure that you are just following basic, you know, good analytic practices of like what's the question I'm trying to answer? And that could be your hypothesis. And then you have, you know, following your uh data requirements, what data do I need? Am I using the right data, sort of following the six C's of data quality to make sure that you can answer the hypothesis either positively or negatively?

It's even simpler than that. Are you following the scientific method? If you go back to high school or even grade school when you first learn the scientific method, you know, you have a hypothesis, you create a test, you validate the results, you prove or disprove the hypothesis. When you're when you're doing hypothesis testing, you are doing science. And it you don't even necessarily need technology to do any of that, right?

Uh as long as you're you're collecting data in these statistically so sound way. I was looking um this past week at some of the early data science that was done by uh WEB Dubois, who was uh one of the first presidents of the N NAACP. This is for Black History Month, and in 1900, he published this enormous register of 66 different charts and analyses and stuff uh about you know life for black Americans in 1900, no Excel, no tableau, no R, no Python, hand-drawn everything, but it was part of this overall hypothesis that you know the effects of slavery and stuff had an impact on black Americans, and he proved it very clearly through this uh 66-page publication, which you can find in the Library of Congress. You don't need technology, you just need the hypothesis and then the scientific method. So even back in 1900, uh somebody like you know, Mr.

Dubois was was just knocking it out of the park. So, what would you say, John? Have you applied data science techniques uh to everyday life for you? I feel like when you um change out some plumbing, you might apply data science. You know, the hypothesis is this thing is leaking because the person who previously installed it wasn't paying attention.

Let's prove it right or wrong. Yeah, well the thing with that though is you know there's zero tolerant tolerance for failure in those tests. So I wouldn't call that true science because yeah, I I would call in a real plumber if I was doing science. We don't want anybody experimenting on their natural gas in their house that could end up very poorly. Um but yeah, it is interesting how it's that experimentation mindset, you know, because that's always been a cornerstone of marketing programs that I've been doing is you know, you're always running two or three things to figure out what works and what's better.

So yeah, it's just kind of funny how um you know, a lot of the data science I've done is about Dubois level of, you know, this is the stuff that was just lying around. We're not using any specialized math or calculus. It's more just okay, let's try three things and see what kind of trouble we get into. Exactly. Now I will say the tools do help.

That you know, the tools, it's like cooking, right? Um, you can do everything with just a frying pan and a spatula. A lot of things will not turn out well, right? And it's a lot of extra work. Like if you're trying to make a steak, that's that's okay.

If you're trying to make soup, it's a little tougher to, you know, you could do it. Um, but having the right tools certainly does help. It's like having a blender, you know, does help if you're making margaritas. Uh it's not essential, but but it it sure is is nice. Um but w the big confusion point that I think a lot of people have with data science is everybody spends so much time on the technology, on the tools, and they don't they ignore the first three parts, which is you know why you're even doing it, um, who's doing it and how are you going to do it?

And without those parts, it gets very messy. Now, in terms of practices in data science, one of the the most important prerequisite practices before you can start doing the science itself, is uh technique or system of techniques called exploratory data analysis. And this is all about given a bunch of data. Do you know whether the data is in fact uh you know usable or not? Do you even know what the data is?

And in marketing, we've got a ton of this kind of data. You got uh you know, data from every conceivable tool. So I figured we'd spend a little bit of time taking a look at what exploratory data analysis looks like. We're gonna use um the programming language R. You can again you can do exploratory data analysis with Excel.

I am not good at Excel. So I will say that you know, you know, someone like Osju Soloy, who's a Microsoft MVP, could probably do everything I'm about to do in Excel. I just happen to do it in R because it's it's packaged up a lot faster for me personally. Um let's start with something straightforward. This is good old fashioned keyword data, right?

This is uh from the SEO tool hrefs. Um, and this is our keyword list. Uh uh here, trust the insights of the search terms that we think are important. You can see there's a lot of stuff here. Um, there's a lot of data in here.

So, what I think our first step should be is let's take a look at this data from an exploratory perspective. I'm gonna run a very quick analysis on it, and we'll get a nice uh set of charts out of this. So it's it won't be you no one we don't have to start at the R interfaces very long. Many of today's modern data science tools can produce these kinds of profiles, these reports that say, here's what's in the box. You gave me this bunch of data, here's what's in the box, and then if you know how to interpret the report, you can get a sense of can I even use this data?

If I can, great, let's proceed. If I can't, you know, maybe maybe this you know we need to find a different data source. So what we've got here from our keyword data is a quick look at the types of data in here. How many things are discrete columns, which means text or uh non-numbers, and how much is continuous, which is numbers. So in this data set, which is our SEO keywords, about 41% of the columns are not numbers, and 56% are numbers.

We don't have any missing columns. When we look start looking at observations, how many rows are completely all the data is in them? 77%. That's a yellow flag. Like we got some missing data in here.

Um, and then we look at missing observations, 3.7%. So just with this alone, Katie, you mentioned the six C's of data quality, right? Clean, complete, uh correct, etc. We're already starting to see there could be some problems in this data set. So we want to keep exploring, keep going through this.

But this would be a first step. Now, if this was like 20% complete, I'd be like, uh, we're gonna this data set might not even be usable. Most tools will tell you know how the data is structured, especially when you get to really complex things, like inside your CRM, like if you have HubSpot or Salesforce.com, you might have 30 or 40 different tables all links together. And tools like exploratory data analysis can help you to start to work out. Okay, there's a lot of things that are interconnected here.

Uh this one again, this is a real simple table, just uh SEO keyword data. Going on down, which data is missing? So now we're looking at the individual variables. So we're good for search volume and country and keyword. Uh we're we've got some missing data on traffic, some on uh search features, last updates, uh keyword difficulty, and we've got a lot of missing data on uh uh pay-per-click data.

So we in immediately that tells me because this is SEO keyword data, we may have a lot of terms that don't have bidding on them, right? There's no there's no auction data. So that could be good or bad. We don't know yet, but we definitely know that there's chunks of data missing. So I guess it really depends on the question you're asking whether or not that's acceptable.

Exactly. Like if we wanted to know what country, then it looks like you have all of your data. Exactly. Let's rewind. Katie, why would we be looking at our our keyword data to begin with?

What's the purpose? Uh the purpose of looking at our SEO keyword data is you know, there could be a few per per per purpose eye purposes. Um the purpose could be, you know, we want to know uh are we ranking for the right keywords, or are there other keyword opportunities? Or you know, um, what are our competitors ranking for, or depending on, you know, we're trying to target a certain part of the country, what keywords do they care about the most? And so there could be a variety of reasons, but it all needs to sort of go back to keyword data in the first place.

Exactly. So for this data set, I think it would be a good to look for essentially keywords that maybe we should be uh trying to find opportunities for. I think that's a pretty pretty straightforward purpose, right? Uh are there keywords that would be a good opportunity for organic. So if this was paid, if we said, well, what are some keywords you can bid on?

Immediately looking at this chart, go, wow, we got a lot of missing data in here. This data set might not fulfill that purpose if there's a lot of missing bid data. Moving on down, we have our distributions. Uh these distributions are essentially show you how the data is broken up for these different variables. So uh keyword difficulty, just looking at this chart uh from zero to 100, we can see that this data set has a little bit of a lean.

You know, if you draw a line down the middle, it leans more towards the right-hand side. That means that from a uh a difficulty perspective, we've picked a whole bunch of keywords, and a lot of them are on the difficult side to rank for organically. Um we look at sort of global search volume. This one leans very heavily to the left side. So we've got all a bunch of keywords that don't have a super huge amount of search volume.

Again, kind of a I call this yellow flag, like so we picked a bunch of keywords that are tough to rank for but don't get a ton of volume. This may not be the best strategy. Yeah, page one. I was crying when I saw that page one. It's actually this chart at least shows it's not as bad as page one indicated there.

I saw that sea of red and was just like, oh my god, this is gonna be ugly. Exactly. Um traffic potential, things like that. So we can see that there is some skewing. So again, being able to look at a distribution, say, does it lean left, does it lean right?

And then what do those leanings mean? The bar chart frequency here is not super helpful. We have uh these sets of quartile quartile plots. This is a very much a statistical thing. What you're looking at here are the first two quartiles of a data set plotted against the second two.

And what that means is you're looking at whether there's a uh a skew, a bad skew in the distribution. A normal distribution should be look like this with a nice horizontal uh uh diagonal line and all the data clusters on it. That means that there's a uh it's sort of like a bell curve essentially, a normal distribution. When you have weird wacky shaped things like this, it means that your data is not normal, which means that that dictates certain types of regression uh techniques. So when we think about the process of data science, we start thinking about what techniques are on the table for us to use.

A diagnostic like this test says, okay, you got to take some things off the table because the data will not permit you to get a valid analysis for certain techniques. So in this case, if I want to do a regression, say like uh for traffic potential, you know, how much traffic could these keywords drive? If I wanted to do regression analysis against that, there's uh three major regression techniques that you learn in stats 101. There's Pearson, Spearman, and Kendall Tau. This weird shaped thing instead of this lovely diagonal line takes Pearson off the table.

So I um, which is important to know because a tool like Microsoft Excel out of the box uses Pearson regression. So if I were to try and do this analysis in Excel, and I didn't do my look at the data uh uh exploration first, I would go, oh, I'll just run a correlation analysis and look, you know, here's this correlation, and the data would actually be saying, uh uh-uh, this won't work. Uh you're gonna come up with an invalid answer. Is that because of the one of the first things that you that you're shown is what's missing from the data? In this case, it's it's that it's a non-normal distribution.

So it's more of like a power law distribution. When we go back up here to these charts, right? Your your traffic potential, like this is what I care about. How much traffic can I earn with my keyword with these keywords? Look how heavily left that skews to one side, right?

There's a whole bunch of keywords in here that have very low traffic potential instead of a bell curve, like this is you know, CPC is more shaped like a bell curve. This is not in any way a bell curve whatsoever. And therefore, that means that I can't run uh uh piercing regression against traffic potential and understand and get a sense of what other variables correlate to it because I can't compare this bell curve with this non-bell curve, right? They're they're fundamentally incompatible. So knowing that I would have to choose a different regression technique.

You may be getting to this, but are there rather than you running um the script that you ran against this SEO data, are there out of the box or third-party services that can do this kind of exploratory data analysis to verify that your data is good or your data is not good? Or is it something that you as the person doing the analysis really need to be doing yourself and building your own set of scripts and code and you know, building it out in Excel? There aren't any so you actually asked a really, really important question. There are not any out of the box tools that do this with marketing data, right? Okay, and that's such an important thing because a lot of marketers will fire up a tool like SEMrush or HRFs or whatever, pull the data and start working on it immediately, not knowing that the data might be bad depending on the purpose, right?

They because nobody ever asks the question, can I even do this? Um then once you run your uh your campaigns or whatever, you may find that well, actually the campaign didn't generate results. Well, why? Nobody can figure out why. One of the underlying causes could be because you use the wrong techniques to draw conclusions to make decisions, and as a result, the decisions you made were in based on incorrect uh interpretation of the data.

Got it. Okay, so there's really no, I can't like you know, import my my CSV file into you know check my data.com or something and get I think I just came up with a great idea. Um, you know, trademark, and get this kind of results to say your data's good or your data's bad. So that for this in this application doesn't really exist. Exactly.

And part of the reason is because it goes back to purpose, right? When you look at we think back to our 5P cycle, just because something is a power law distribution or a normal distribution doesn't make it good or bad. It is a question of how are we going to try to apply it, right? It's just like a blender isn't good or bad. It's the context we use it in.

Blender for margarita, great blender for steak, no, not so much, right? The blender itself is not the the issue. The same thing is true of our data here. The distributions matter based on the techniques we're going to apply to them. Um a tool, uh a third party piece of software would not know what our stated intent was.

I see. Well, and I think okay, that's super helpful because I'm sure there are tools out there that can validate your data, but unless you're telling it why you're looking to validate the data, which you know, these are not as far as I know, you know, sentient things, you need to put that in. And so that then goes into the whole programming piece, so you might as well just do it yourself. Exactly. For now, we are seeing more and more tools come on the market that do more data prep and try to make it easier on people.

You know, for example, Tableau released in the Tableau Data Prep tool. There's been tools like Alteryx, IBM Watson Studio and and uh data refinery and stuff uh that are out there that can do bits and pieces of it. But again, goes back to the purpose. A lot of these tools are general purpose tools, kind of like a Swiss Army knife. And if you don't know what you're gonna be doing with the tools, then the tools obviously can't conform to those specific use cases.

So you in a lot of cases, yeah, you are gonna have to you may have buttons in those tools that can perform these individual analyses, but you still have to understand why you're asking these questions. I feel like this is where John busts out the duct tape and you know starts putting tools together to faking things. This is actually though, horribly disappointing, really, when you think about it, because we know that one third of you know, marketing orgs aren't even using data at all. So now you have the two-thirds that are, at least a third of them are making completely erroneous assumptions. You know, that they have this small little bucket of data and are just presuming that it's gonna be normally distributed, so that if they throw in money, they're gonna get there.

And yeah, unfortunately, for over a third of those, they're just it's not going to be a fun lesson. It's going to be crashing the you know the bus into the side of the building at full speed. I mean, it's you know, it's not gonna work. So um, yeah, I guess do your due diligence and look into the data before you promise uh quarterly results because that could end very poorly. Exactly.

The next thing that this particular tool does is does a correlation analysis, right? And again, this is where that gotcha comes into play. Let's say I care about traffic potential, right? I want to how much traffic can I get? What variables correlate most with traffic?

We go looking across the table here quickly, uh, global volume, which is search volume, uh correlates most strongly followed by difficulty, which is is interesting. But because we just talked about comparing a power law distribution to a normal distribution, we know that this difficulty scroll back up here, right? Difficulty is a bell curve, traffic potential is a power law curve. So this conclusion is in this correlation analysis is wrong because uh out of the box this uses a Pearson correlation. So I would have to go back into my code and make an adjustment to use spearman, which is is best for uh non-parametric distributions.

So that's a case where we if you know the underlying data and you know what techniques can be applied to it, you know out of the box this tool needs some adjustment, right, in order to draw a correct conclusion about what actually would be the indicator for uh keywords to choose. Um so understanding that to watch this episode, you have to have some sort of a basic level of understanding of statistics. Um in general I do, but you just said you know, spherman versus Pearson, which are two techniques I understand, but you said a non-parametic distribution. Oh, sorry. Um so a non-parametric uh distribution means yeah, means uh it's a not it's not a bell curve.

Oh, okay. So okay. So a bell curve versus a non-bell curve, got it. Exactly. Okay, exactly.

So we know from our distribution charts, we got a mix of half and half. And so this default, which is built into the tool, um, is only assuming everything's a bell curve by the technique it's using. And again, that goes right back to what we're talking about. Like, why don't tools do this out of the box? Because literally it had no idea the different distributions.

Got it. Okay. And it's horrible because it's like you've got a point three, so it's like it's not even that good. And then you're coming back and say, oh, by the way, actually that's way overstated. It's not even that good.

Right. Well, we don't know. That's the thing is when you're comparing um distributions uh with uh you know, regression different recognition techniques, you don't know. Sometimes they'll come out stronger, sometimes they'll come out weaker, sometimes they'll just give up. Um the last thing that your average tool is going to do is it's gonna do what's called principal component analysis, where it's going to try to figure out what are the most important variables uh in a data set just to try and boil things down and make it easier.

Um, so we see uh analysis number one uh here is uh is at the uh 39% variance, which is obviously in this case better. Uh and so the question is what variables make up component one? It's uh number and then uh CP cost per uh what is CPS in cost for something. Um I don't remember what it is in href's definition. Now, knowing the data set, I know I've got a problem here.

One of the things that href does Hrefs does, which I really don't like, is that in the data they put a row number, right? That row number is not helpful for anything, and so this our software out of the box didn't understand that that really shouldn't even be in there, right? And as a result, it's cut it's made an analysis that's flawed, right? So, and it's again it's not the fault of the tool. The tool had no idea that there's a row number.

It is us as the marketers with our data to look at and go, uh, actually, let's take that out first because that's stupid. You know, it's which is interesting because I feel like you have to do data cleaning to do this validation to figure out if you need to clean the data. And it just like it's this is the kind of thing where like my mind starts to be like that monkey with the symbols of like okay, all right, I think I'm following along. But that one to me is that's a tough one because you do you have to structure the data and clean it first before you can validate it. And once you validate it, you can see how much extra cleaning you would need to do, or if you can even use it.

Exactly. And again, this is why in marketing data science, that process part is so important. You what you're describing is actually agile methodology idology for marketing data, right? For marketing data science, which again, if you're familiar with agile, like, oh, that makes total sense. I have to iterate with my data cleaning until I get to clean data.

It's not a one and done thing, which again, uh yet another reason why there isn't a tool off the shelf that you can just plug your data in and it just does a thing. No tool in the market's going to know, oh, well, for keyword data, make sure there's no the the vendor didn't include a row number. Like with Google Analytics, when you export data out of Google Analytics, you typically get a you know a long CSV file with two different tables in the same file, and you have to know going into it, I gotta delete a bunch of stuff to even make it work in in another tool. Google Analytics also always adds a summary row at the end of its exports, which annoys the daylights out of me because it then throws off everything. It's like, oh, on this day you have five 400,000 visitors.

Like, no, that's the entire year. I don't want that in there. So to your point, Katie, there's a bunch of process stuff that has to happen in order for this stuff to work. Well, and I think you bring up a good point. And agile in and of itself could be a whole episode.

Um, but applying agile methodology to what seems like a cumbersome process, is you know, it's hard to reflect uh on a slide to say like this is agile within this, you know, whole framework, um, because that would be sort of unwieldy with a bunch of you know circles. But that's exactly what you need to do to approach these things is um, you know, if you're looking at it from a waterfall perspective, so waterfall in project management is step one, complete, and then you can move on to step two, complete, and then you can move on to step three, complete. Whereas agile has a lot more things running concurrently. Um by applying an agile methodology to the management of your exploratory data analysis, combining that with data quality, that is where you're going to get to results faster, and that's where you will find out okay, this is data that I can use. Now, with that, it still takes that planning up front to say what is the purpose, who are the people, the process is going to be agile using EDA and data quality.

The platform is going to be this software and this kind of data, and then this is my performance, my measure of success. So you can't skip the planning process. That's where you determine this is how agile is going to work as I'm combining these two things. Um, you know, so that's sort of just a little bit of deviation into how agile would work in this instance. Exactly.

Uh, so that's one example. Let's look at another example. Uh I'm going to in this case uh pull out some data. So I use a website scraping tool called Scrutiny. Uh, and I scraped the entire Trust Insights website.

Let's take a quick look at the data itself. And what we see is whether a page was good or bad, uh, the page, uh, if there's a redirect the the text of the link linking pages together, what page a link appears on, and then when I crawl this, and this is uh uh we we use scrapers, uh SEO scraping tools to be able to understand, like you know, for example, uh what pages are accessible on our website or not. But again, going back to kind of where we started, if we start with a basic uh exploratory data analysis to understand this data set to get a better sense of well, what did this tool give us? Um, we can start to understand what we can do with the data. So we're not gonna spend as much time on this one.

We can see again, we've got the this case 100% discrete columns. There's no numbers at all in this data set isn't that weird there's not a single number it is all just categorical data which means immediately if there's an objective that we we care about we know this data set will not answer that question right so if we're trying to understand for example which pages should link to other pages on the Trust Insights website for maximum say uh moving traffic around there's no traffic data in this thing so we'd have to we automatically know just from looking at this we if we have numbers we care about we need to supplement this data we need to go into Google Analytics and export our page views for example uh but looking right at this otherwise this is actually a fairly healthy looking data set um we have um 89% of pages are missing redirects that's actually a good thing because it means that we're not bouncing visitors all over our website uh on a concern one thing is there's five percent of pages here that don't have link text so we've made a we've put a link uh to another page on our website and there's no text to click on it's like so what did we do wrong there yeah that's I don't I don't know how that works so I I know what that's from but I only know that because I you know uh it's it it's intentional uh it's part of our newsletters um but it's something that if this wasn't our website or we didn't know our data and how we were publishing stuff this would be an element of concern um we can look at the uh error code so most of our pages on our website work fine the a bunch have redirects there are a few that were not checked. Um, critically, no 404s, means no missing pages. I'm very happy about that. Uh, and and nothing weird.

So uh life is good there. The correlation analysis principle component analyses are not helpful because these are all this are all categorical data, there's no numbers. So, even in a very simple examination of uh website crawling data, we can look at the data set and and go, okay, well, here's some things that we would need. And we don't have the data in here because we didn't clearly define that in the purpose. We did say we want to crawl this data for the purposes of you know, figuring out where to link pages uh uh traffic-wise.

Because we didn't define that, we just put data in, we got an analysis that really doesn't fulfill the need. So having that purpose section up front is still super important. So I apologize if your brain is doing the mock arena. It definitely is. It's it's definitely like vacillating between a macarena and the you know, symbol monkey.

I mean, to the point where I'm like, words, what, who? Um, you know, it's it's interesting because I look back at the statistics courses that I've taken, um, and I think I've shared this uh small anecdote with both of you, is when I was in grad school and taking a statistics statistics course, I didn't really even understand the basics of how to make a chart in Excel. Uh, let alone, you know, my professor wasn't necessarily teaching me uh Pearson and Spearman and those kinds of things. Those are techniques and uh methodologies that I learned on the job when I was working uh in the academic setting. And then Chris, you have continued, you know, to educate me on what those are.

And so I feel I do feel like marketers, you know, depending on their educational background and their work experience could be at a disadvantage for some of you know, for general data analysis. I mean, John, you were mentioning, you know, people, you know, you can't, you know, one third of marketers, you know, are using data, and two-thirds of that one-third, you know, are maybe doing it wrong and promising the wrong things. And it occurs to me that, you know, a lot of our clients who come to us and say, well, you know, I put the social media metrics together, but it didn't demonstrate the thing, what went wrong. I think, Chris, you've just spent, you know, 35 minutes demonstrating why the analysis goes wrong and why you don't get the results that you're hoping to get. It's not because you put together, you know, bad campaigns or that you did the wrong kind of targeting or that your copy was incorrect.

The data doesn't hold up to give you the answer that you're looking for. Exactly. What we just did is what you do at the drive-thru line when at Wendy's, right? When the you get the bag handed to you. You open the bag, you look in the bag and go, okay, is that my order or not?

Like, oh, like I guess five cheeseburgers. I ordered chicken nuggets, right? You hand it back to them. You don't have to do this exploratory process all the time. It's when you get a new data set to work with where you get fresh data to work with.

You run it once just as a quick health check. Like, you know, again, fast food example. You don't have to keep opening the bag as you're eating your meal. Once you've figured out that that is in fact the meal that you order, then you just eat your meal. Um, but if you open the bag and it's all like all hamsters, you're like, what happened?

Hand it back, so it just gives you a lot of things. What kind of drive-throughs do you go to? And the same is true with the exploratory data analysis techniques. You don't have to keep rerunning them all the time. This is a health check on the way in.

So when you kick off a new client, when you kick off a new project, when you start working with a new tool, um, this is a great thing to ask people who you know, one of the things that has happened in recent years, uh, much to this dismay of folks who've been in in the space for a while, is there's a whole bunch of folks who you know have taken the six-week crash course and you know becoming a data scientist or whatever, just like you know, six-week crash course and becoming a dentist, like, yeah. Um these techniques are a great way to assess where somebody else is. So if you're hiring a data scientist or you're hiring a data science agency or an analytics agency, these are questions that you might want to ask and say, tell me about your exploratory data analysis process. Tell me about how you uh evaluate a new tool. Tell me how you think about a data set that you didn't generate.

How do you analyze it? And likewise, when we start doing stuff, for example, like when uh I'm finishing up a paper on TikTok right now. The first thing I did when I got that data set out of TikTok is stuck it in here and said, What is it, what kind of health is the data in? What does it have to say? So, John, I know one of the things that you do uh speci uh you know, through you know, the traction model is you like to do those A B tests, experimenting, you know, which channels are gonna be most effective for you.

You know, what do you look for when uh helping a client, you know? Do they have the data? Do they have the right things? Because I would imagine a lot of what we just talked about applies to those projects. You're you're doing that exploratory data analysis, whether it's formally or informally.

Yeah, and lack of data is always the biggest problem, especially with smaller ventures. You know, they just they have some data and they're taking a guess. But it's interesting too. I kind of I think there's a bigger thing there of how so much of what you were talking about, Katie with statistics, as you're starting out with that, you kind of always are going with normal distribution. Like that is the understanding of how things work.

And the reality of marketing is it's almost always power curve that you know there's only three or four winners in a space. And so you just have this fundamental mismatch of all the tools out there are you know working in normal distribution world, where the reality is, you know, the marketing team only needs to find one program that works. And if they find the one that works, it's gonna crush everything. So yeah, again, it's um just keep panning for gold, you know, use the tools to try and direct you in the right direction. Um, but it is still kind of an all or nothing thing.

Uh you know, I'd love to say that you're gonna test 10 things and four of them are gonna work average, but it usually works out that you're gonna try 10 and you're either gonna destroy it with one or go out of business. Um, the other thing that occurs to me, uh, Chris, as you were going through all of this, one of the questions that we get a lot, um, probably in every single conversation about any kind of coding, data science, machine learning is will the machines take my job? And I think you just very clearly demonstrated the answer is no, because there was so much human intervention needed in order to tell the basic code, the machine, you know, the output, what it is that you needed, and you have to keep going back and doing that and revising it. That part of that agile methodology of you, as the human need to determine the purpose, and then you as the human need to evaluate the data set to say, are there row numbers? Is there a summary, you know, row?

Are there you know columns that are strings and they should be this instead? Like, unless you, the person tells the machine that's what's supposed to happen, it doesn't know. It's not gonna guess at it. And so I think you've just made a really good uh point as to why, at least in this context, humans are very much still needed to run the machines. I was just the machines aren't gonna take your job.

It's gonna be a person who's better at using the machines than you are. Yes. I would agree with that. When you think about that question, the the answer logically is will appliances take the chef's job. No.

Someone still has to put the tequila and the limes in the blender and push the start button. Maybe it automatically senses that there's stuff in there and it starts on his own, but someone still needs to get the tequila. So that that is that is going to be the answer for the foreseeable future is no, the machines are not going to take away the jobs. They will take away tasks. Uh, they will do individual tasks better.

And you overall may need fewer people to get the same number of tasks done, but a machine is not going to take your entire job. I mean, uh there's just no way for a machine to be that multidisciplinary. All right. Alright, sentient margarita machine. There's something there that can be.

Yeah. That could be a we'll do the sentient margarita machine and the check my data.com. Those are the two actions that came out of this episode. All right. I think that's enough for this week.

Thanks for tuning in, everyone. Thanks for watching today. Be sure to subscribe to our show wherever you're watching it. For more resources and to learn more, check out the Trust Insights Podcast at TrustInsights.ai slash TI podcast. And a weekly email newsletter at TrustInsights.ai slash newsletter.

Got questions about what you saw in today's episode? Join our free Analytics for Marketers Slack group at TrustInsights.ai slash analytics for marketers. See you next time.