Hi everyone, my name is Yana Lyashenko. I am a Google logistician. And in today’s video I’m going to show one of the team’s case studies, but I think this case study will be of interest to more of those nerdy nerds who like to dig into analytics and understand why Performance works this way or that way. Or savvy nerds might find some new insights for themselves. It will be a completely different case format, I won’t show you what an awesome result we’ve made. And how else, let’s do it this way, argumentatively explain that those things that the specialist does in the account – they are not just a test to some random hypothesis, but they have a basis for themselves.
Plus it will be another account that will prove that not everything in the accounts is random. Because in the auto parts industry, some experts believe that the results that are made are random. In those who have a great it happens and mostly merit – it’s just seasonality. Here we will show how it all correlates, depends.
The period that’s being considered
So I want to show, emphasize the period, will there basically compare this fifth somewhere before here period. This is just the period when it more or less settled down after the first tests. We have a start date of the 17th, this is the moment when our new campaigns were included, the introduction of basic adjustments – this is the date from which maintenance began, the 17th.
What purchases were recorded for statistics
Here we have purchases recorded, and it is quite interesting what kind of purchases. Before that, before our period, we mainly recorded the fourth analytics, purchase, which included two goals: ordering through the cart and buying in one click. Because there is a good percentage of orders there and there, so it combined these conversions.
But I’m going to show exactly in this video exactly what the downside is of using this conversion fourth analytics, why it’s still worse than tag buy. And that in the end even on these modified sort of attribution models, how this even on purchase with fourth is positively impacted. Because then we have separately configured shopping through cart, one-click shopping specifically to separate them, to capture them separately. RVC is available for both of them (RVC is advanced conversion tracking).
A brief overview of the client and maintenance periods
What are we doing here? I’m going to show you all of this now so you don’t think I’m drawing something in numbers. This is the total price for conversions of 2.76K for this whole period. We can see this period is more or less stable for the client. Right? We can say that somewhere up to this point even the conversion did not fail. Here you can see that the main average target price per conversion is 136, 113, 125, 100, sometimes hryvnias. This is where the client tried to do something on his own. There was a complete mess going on. And here’s the launch date.
This is where we tried the launch. You can see that the price per conversion is about similar to this. There are slightly more conversions than there were in these initial periods, in pieces. We made a certain shot, which we managed ineffectively, as you can see here. There were 45 of the same purchase here, which, by the way, were also 44 here, but it is much cheaper here now. There is also a shot here, and everything is equalized here now, everything will be fine here.
By the way, these perches, seals that you see here now 50 – this is exactly the same number of equal conversions, not these half conversions that we had before. These are half conversions not of the purchase tag, but of the fourth analytics. Now I will explain what the problem was with these half-hearted conversions.
The effect of the fourth analytics
I will show the basic analysis on the effect of this fourth analytics, because this is the only goal that existed from month to month. I have it recorded in the “All conversions” column. Periods, here look – 15.05-16.06, here 17.06-19.07. 19.07 even I should not, apparently, take. 19 is a period that hasn’t been finalized yet. Here’s a look at that whole story, now just a moment.
Yeah, somewhere, “All Conversions” is ours. I have this purchase here, so we will base ourselves on it. I will show you that the growth here is much higher, maybe even more, and even the price per conversion is much cheaper here. Where’s our shopping? One-click shopping, it’s been redone, by the way. It’s corrected to work correctly. Well and the Buy Tag that worked here.
So if you add them up, the approximate sum may be a little more than that, but in reality the client had a completely different number of orders. That is, there were more dirty conversions, and there were somewhere up to 1,500, something like that. That’s up to July 14th. And this is me showing up to the 18th. This is me showing in general how it all affects the picture. I’ll explain now.
Setting up advanced, tag buying, stabilizing prigans
So, what was the task? First of all, to set up the extended, Tag Purchase, so that there are correct transfers, to give more stuffing to the Performance itself, to stabilize the history of this ducking. Because this attraction in the results started before our work. I mean, it’s a conscious effort, right?
In order to get through these things. In general, this is all the inducement here, it’s not due to anything correct. Even here there are 12 conversions. The price per conversion went up very high and such prigans started here.
What’s the cause of these priggans? There could be quite a few of them. Insufficient, although there are a lot of clicks here, machine modeling, which does not work correctly on this particular project. Perhaps there is also the fourth analytics itself, it is now constantly trying to be improved by Google and it can fail. Plus it’s not quite correct for most accounts in terms of tracking and scaling results. How can I verify that this is the case?
I generated a report like this in the report editor. I took just “Action Conversion”, the Item ID of the item itself, I broke it down by day, I took all the conversions in order to filter out only perches. This is the purpose of the fourth analytics just purchase here, Item ID and here is the breakdown by day, from May 15th something like that. Here I have a period of equalization here 17.06-18.07 I’m equalizing 15.05-14.06. This is about the period I just took for equalization. This is by purchase, I want to show by purchase, because by purchase it will not be seen, but I will show how much it, even, positive the very work of advertising campaigns affects in general. What the same fourth analytics doesn’t notice, all these crosstabs. What Tag buy is noticing now is that it’s a much better kind of conversion for Performances and it gives you more analytics.
Right now, the only thing that the client should be able to do is….. I mean, we’ve only had a month of the sapporter. In the month, we tried to get a little more aggressive in pushing through all these things. They are not always successful, but the longer you work, the more stable the result. These fluctuations that occur, the jumps – they will happen regularly. Because you have to understand – Google’s advertising tool is not static, it’s constantly changing. And this “constantly changing” is not once every six months, or once a year. It’s just very regular changes. They just change their merchant verification policy once a quarter, maybe even more often. I wouldn’t be surprised if it’s once a month. What’s to say about these tools? Especially since there’s an A.I. chip built in.
What is the complexity of the attribution model?
What do I want to point out? The basic complexity of the attribution model that fourth analytics uses, even if there… By the way, on July 12, the “last click” attribution model was exposed for paid channels only. It is still not enough for Performance to work correctly enough on the fourth analytics. Only Performances worked here. Because look, I’m taking the previous period. I want to see all the more or less reasonable orders. Orders I mean, I want you to see in reality how many full-fledged orders Performance itself can make. Of course, this is not a sufficiently correct representation, because half conversions are still present here.
That is, this figure that I have filtered out now, I just took “more than one conversion” by IDs. But this is the report that you can very easily generate and check. These are the so-called “non-longtails”. Merchandise items make more than one conversion in this period, the old one before the previous one. What do we see here? What we’re seeing here is an average of… That’s 725.56. Of course, there’s more halves here, there’s a lot of them. You see, there’s a lot of them. But we filtered out those positions that made more than one conversion.
If we remove the filter now, the system will show what? 887 conversions. That’s 725 and 887. Come on, 887. It’s just that even without that 725, that’s 162 conversions. Here he added 162 conversions to himself as a positive result of some kind. It’s the same story here, but in these periods the campaigns worked on completely different mechanics of the scaling and adjustment procedure model. How was it introduced? What was the goal? To try to scale within the KPIs that are there what was there. Plus increase some things. But at the same time to make larger figures on which to build.
For example, I will take the same filter, but in the period when we were already working with advertising campaigns in a new way. I’ll take the same “greater than one”, same purchase target. Based on the data here, 1149, even with halves.
Same selection principle with the same halves, although there may be significantly fewer halves here, as you can see. Now compare the 725 we had, right? 725 divided by 1145, that’s 63%, minus one. That’s 36.6% just growth. 36.6% even with those halves, but product lines that are more stable in subsequent periods. We can already see this on the attribution model based on the data that’s in the advertising account here on the fourth analytics. If we look here, it’s very hard to see.
So we have an increase, if you look here, of only 27%, but it is actually much higher. I mean, what’s my count there? 36,6%. Let’s say 35%, I’d like to round it up to 37%, but let it be 35%. That being said, we can roughly even calculate…. I’ve got 15.05-14.06 here, right? 15-е? Well here 16.05-16.06. Let’s equalize these periods as it was already here I counted. 17.06-18.07. Let’s put 17.06-18.07 and what was the period here? 15.05-14.06. Let me put this 15.05-14.06 here as well. 15, just one day off. Two days, even. That’s right.
Comparison of the result in price per conversion
Now let’s calculate what we actually got in price per conversion. I’m gonna find our bones in here somewhere. Bones, the conversion’s up here. Yes, conversions are up. And here are our bones. Let’s now calculate how much this conversion price cost us in total. Look, before that we had 104447.42 UAH.
Expenses are not counted as if by the attribution model. You agree that models don’t count that way. What do we have here? 104447 divided by 725, rounded up. This is 144 UAH price per conversion. Fixed – 144 UAH. And now here we divide 135757… How much did we have in there? 1149? But I’ll double-check just in case. 1149. Divide by 1149. This is our price per conversion – 118 UAH. If we divide 118 by 144, it turns out…. Minus 1 is an additional minus 18% minus the conversion price.
I analyzed this directly by perchase. That is, without taking into account the fact that in previous periods the Buy Tag was not fixed. So, accordingly, it is very difficult to show it. We can say that here there was an increase of 35% in the number of conversions, even those half conversions. Because we also measure the last period with halves, but even the price per conversion became much lower.
I realize I’m showing a very super primitive analysis. Somewhere, maybe not exactly equal numbers that I would like to see, but this is also a cool result. This is taking into account, for a minute, these shootings, when the price for conversion went up, was quite high, and then it came down. Here it is 109 hryvnias, almost like it was here. But here again this period it has stabilized. And here in this period, by the way, from July 5 to July 6, here the price per conversion is more or less stable, and there are more whole numbers of conversions. So you see, 39, 40, there were some halves here, that’s our attribution model. Gosh, one-click shopping. But here, it’s basically 36, 50, here for example, here’s number 18, 50 pieces.
What is the flaw in this fourth analytics?
What is the disadvantage of this fourth analysis, even though I’m analyzing on this fourth analysis. And it will be in the following, what is it? I want to show you the next point that the fourth analytics may not take into account, and what will be assigned to itself, for example, the same Google purchase, and what will it show us? For this we will generate, for example, a report. I like it very much, I will remove the “Product Title” to make it more convenient. That’s it for the crosshairs, this report is ready for you.
That is how you can get it, you go into your report editor, and you will immediately have a block in this block, which will be called “Shopping items sold” something like that. Click on it, it will be the same. “Product Title” is out.
What can we see here? We can see here, by the way, what the fourth analysis will not see. Unfortunately, I will not be able to show you the previous period before the start of work. Why not? Because at that moment I did not have the icons highlighted, not shown. So now I will demonstrate it to you so that you can understand why I cannot compare this report by previous periods. If I select a previous period, the product identifiers are not recorded here, they are not transferred. This is a common report in analytics, in the online store purchase reports. I will now show you the difference between them.
Here is the last period, here it is recorded – 21.06-18.07. What I want to show you here is what Google itself does not take into account, for example. We have cross-assignment positions. Right? I mean, we can probably find some or half conversions here, right? For example, a filter greater than zero. We can now take the same conversions on perchase, right? I’m going to show you what’s going to happen here.
Okay, let’s see, we have a product item under this ID. It did both direct sales, and it did the same cross-sell. Or even better, find a position that made only cross-sells. The Google system will take into account this crosseil, record it, fix it and use it in optimization.
That is, here the sale was directly on this product item. If I turn to analytics and try to find this item ID 14179 for this entire period, I will not see any conversions on this item at all.
See? It wasn’t acquired here at all under the attribution model. What I’m going to take in this analytics report, I’m going to take out all the filters. I’m gonna take this very ID number, see if I can find 14179. I’m going to find it now, apply all the history. And I won’t see even half a conversion here either. And Google purchase sees that here was a crosseil, that is, this product…
What do you mean this crosseil? That product item sold by click through an entirely different product. That’s the coolness of exactly the Buy Tag even on the Others attribution model not based on data, but compared to even similar settings for the fourth analytics. So the fourth analytics doesn’t see it at all. Yes, it saw, maybe there in June, June 12, some 0.5 conversions. But this one, for example, conversion… We can try it. It’s between 22.06-19.07. We can even add a “Day” parameter such as. Row – Day. Take find this product item, apply and see that there was a crosseil at all on June 26th. And it’s not in the analytics.
Why is it necessary to do the right work in the account?
Because when everyone talks about moving from the fourth analytics, work with the Buy Tag for that reason. But that also wouldn’t have happened if you weren’t doing the right work in the ad account. I don’t know, maybe it sounds very pompous or something like that, but I can’t help but emphasize that the result is done by the hands of the team. With the hands of the team is made this result, here the adjustment of all the changes, the structure, excluded, filtering traffic is made, excluded, my structure is made (where the video, it seems that it is a chatter on the structure of Google shopping) – it is taken into account. So a bunch of different factors, including the right settings.
The key here is in understanding what you are going to stuff that Performance with. It seems that here the result is not properly justified at all. The base starts, then some slippage, another slippage, and then a certain picture emerges. It is hard not to notice that there is a positive dynamic here, right? It’s hard to miss. Plus, if we’ve seen half conversions here, right? Let’s take July 5-18 when we set up, conversions were fixed. And, for example, let’s take the month of May. Let the fifth be the same as the 18th. And I will also equalize with June. Equalization 5-18 is a period of ex post facto adjustments to stabilize. What do we have here? Some halves are smaller, right? Some halves less. The cost per conversion is less. We can do the same analysis here, which I suggest you do so that we can see comparative analytics. So that you can see objectively that this is not some fiction, for example, mine.
We take again the period 5-18 and here we equalize, for example, 5-18. I take similar dates to make it more convenient. See here, it’s already on perchase. And to be fair, right? Without the Purchase Tag, but so that we calculate using the old models, because here’s the old conversion rate. We take “less than/equal to one.” We’re looking at what? 410, even with the same halves.
The filter is removed, we take the same filter. I’m showing you this step-by-step so that if anything, you can do it in your advertising account.
There’s 465 pieces, I’m taking that out. 465, 487, even though the full is 410. So if I take the “less than zero” filter now, even take “less than/equal to 0.99”. See these 0.5, 0.02, 0.3 conversions right here? They generate 72 full sales. We actually have 410.
Then we go and look at the expenditure side, right? So we take the same period, for example, that’s what we had? 5-18, right? 5-18 compares to May. Why am I taking May? Because it didn’t have all the overhit stuff we had at the start. 5-18 equalizer. And now let’s take a look at the bones here, 410, 465. Let’s do a simple comparison.
Here the expenses were in the previous 54,000 divided by 410, we get…. Now, wait a minute. 54472 divided by 410 – we get 132 UAH price for the sale. Dirty, most likely it is even more expensive, because there are halves there. And now less costs for the same period 43893 divided by 465, also with halves – already 94 UAH.
Analyzing the results of adjustments
Even if you compare this period of relative stabilization, after all the parts and adjustments in four weeks, it is very difficult to make a very sharp jump very quickly. When you see enough analytics, you can try to break through it, but it does not always end with a result. We have clicks ducked right here as we saw, right?
If we take the 5th of May we can see that here we have clicks ducked, of course the cost of the price per conversion has increased and we had to roll back certain parameters and a little bit of patience to stabilize and make very different adjustments in the period of stabilization of advertising campaigns. But you can also see here that even a relatively stable period right here, right? I wouldn’t say I took it straight. I didn’t take the 19th when the price for conversion went up. I plus or minus took a more or less stable piece to analyze and it turned out that there 132 UAH. taking into account more halves, right? And here it turned out to be 94 hryvnias. That is also much less.
If we take even these 94 UAH divided by 132 – how much is that? Also by 30%, even less price per conversion. That is, when I analyzed the total period there, from the 17th or from what date to the previous four weeks – we had a decrease of 18% price per conversion. Here we have even 30% more or less stable period. Even with these at 109, at 139 UAH price per conversion. With this interval even that more or less looks stable. I directly compared to what was good, not bad.
Conclusions
The purpose of this case study was to demonstrate that there are no accidents here, that there are some things that do not depend on a specialist. Of course, seasonality or demand for some things – it positively or negatively affects the overall dynamics. But here we have plus or minus, it can be stable. You do not see here that somewhere the conversions have completely dropped, fallen. It is a smoother set of masses somehow here it became more conversions and so on. But it’s an equal number of conversions to see it in equal proportions, equal numbers, for example, take the same period from the 12th to the 18th. Here it’s good, here it’s not quite equal because here it’s a one-click purchase, it’s left to the data-driven attribution model. Here it’s 25.31. That’s even right there.
Let’s take it from the 13th, a partial week. You can see 202 conversions here, they’re perches. Who are the non-believers out there, whether they are or are not perches. Here’s the convincers/perches. This is the same type history and please 98.19.
This is an example of a period of relative stabilization, what can happen in an account when a specialist knows approximately on the basis of this analytics, which he does not quite quickly see, right? We cannot analyze the conversions for today correctly, adequately, because for today they may be much less and when we look at yesterday and today the day after tomorrow there will be a completely different, more presentable number of conversions. He knows how to operate with this.
