These fields act as a tag, and for every promotion that you create (from a specific list), you will choose to which tag of customers the promotion will be created.

You should create as many fields as promotions you want to have active at the same time. Then, when you're creating a promotion, you'll choose between the tags you created, so we can tag that group of customers, and direct the promotion to that tag.

We recommend the tags to have generic names, because you want to be using them repeatedly, for different lists and different promotions.

Only if there's a list that is of particular interest to you, and you think you'll be creating promotions for this list regularly, it makes sense to create a field specifically for that list.

We recommend to create at least 6 fields, as it takes some time for VTEX to update and reindex your customer catalog every time you make changes to it, such as adding or removing fields.

Creating fields

Ok, so hands on to create those so-talked-about fields.

1. Log in to your VTEX Master Data admin panel, and under Aplicações, click on Advanced settings.

2. Click on Estrutura de dados (data structure). You will be redirected to a new window

3. On the new window, click on Data Entities, and find de Cliente (customer) data entity. The Acronym is CL. Click on the edit button at the far left.

4. You will see all the fields that already belong to your Client structure. Leave all of them alone, and create there the new fields. You will have to set 3 things for them: Name (for us to identify the field. Please don't use special characters not spaces, only numbers and letters), Display Name (here you can put a short description of the field and use any characters you want), and Type (this is the kind of data that will be stored in the field: it can be numbers, letters, etc. We will ask you to choose the type "boolean", which is a data type that only support "true" or "false" as a value.

5. Once you've created all your fields, click on Save. You will be redirected to the previous screen, where you saw all the data types.

6. Click on save. You will be asked to confirm you want to publish the changes, as they will go straight to production. The message also warns you about not deleting fields. Make sure when you click this button, you have NOT changed any other fields, but only created new ones.

7. Now that we have published the changes, a text dialog will appear saying that if you want these fields to be searchable (Buscável) or a filter (Filtro) you need to click the re-index button.

As we WILL be filtering through these fields, we need to do this. So close the dialog and then click on the reindex button for the Client data entity. This is the 4th button on the right, with 4 green arrows pointing outwards from the center

8. Done! You will get a message that confirms the system is reindexing on the background

Now you're ready to create promotions specifically for the fields you created! What are you waiting to start working on your next marketing campaign?

We present list exportation directly to VTEX promotions. Promotions are a feature VTEX offers to their customers, so they can configure very specific promotions to be applied to their customers, allowing segmentation via numerous fields, such as email, birthday, previous purchases, order value, specific products, etc.

Even on their own, these promotions are a great way to increase sales, create and reward loyalty from customers, and help with customer retention.

However, unite them to the power of Sensai and you have an unstoppable tool!

Why? Two main reasons:

1. CLV segmentation

You can now send a promotion to a specific cluster of customers, based on the AI-predicted CLV model we have. This way you can send a promotion to your top users to incentivize the sale of a new product, or to get them to bundle up that purchase they were gonna make, increasing their AOV.

You could also send a promotion to customers who are at risk of leaving you, to make them happy and give them a reason to stay.

Or you could send your special deals to your safe customers, to make them feel special and reward their loyalty.

There are many possibilities, make sure you explore all of them!

2. Custom list segmentation

This is the star of the show, 'cause now you can send a promotion directly on your store's platform only to users who have taken special actions that you decide in the creation of the list: having bought that coffee machine that needs these capsules that you're just pushing to the market; having spent over $1,000; having purchased a household item but also a fashion accessory, etc.

The way these promotions work with our lists is we take one of the lists you already know and love from our platform, and then tag your customers in the VTEX platform to let the system know those are customers that will receive the promotion. Then you create the promotion, segment it the way you want (to apply to all products, to specific categories, to certain merchants, etc), and with a couple of clicks, you have a new marketing campaign, tailor made, super segmented, active (or scheduled) on your site!

Try the promotion tool now from our app, in the CLV section!

As you know, the balance between an e-commerce's metrics is very delicate. Not only are there lots of measurable things, but each one affects and is affected by many others. Determining the budget for marketing, and deciding the division between channels is no easy task, and requires an understanding of these relations.

However, as there are so many variables, and some relations so delicate it's even hard to notice them when analyzing data, we've decided to build a tool that puts together experts' knowledge along with Machine Learning algorithms, to give you the best budget allocation model you can get.

The model

Because of all the involved metrics, and the importance of your budget allocation, we provide you with a wide range of information from this model:  of course, the budget you need and its distribution per source, but also the number of sessions you need from every source, including organic sources; the size your audiences should be for you to be able to bring those sessions; the revenue we expect you to make; and how much of that revenue is going to come from returning customers, and how much from new ones.

We follow a number of steps to estimate all of this, but it all starts with the revenue you're gonna make.

a. Revenue

The first step for us is to predict how much money you will make, following the tendency you have had so far.

In this first part, what we do is an accurate prediction of how much revenue you are going to make over the next 30 days. This prediction considers not only the seasonalities your store has, but also the special dates (like Black Friday), and lets out outliers (like really specific campaigns you had for one or a couple of days that shifted considerably your revenue of the period, or B2B sales).

Now, our model is quite special, because not only do we let you see how much money you're supposed to make, but we also let you plat with your numbers a little. After we give you the estimated revenue, you can choose if you wan to 'keep' that prediction, or if you actually want to make a little bit more or a little bit less money. You can adjust revenue by adding up to 30%, or by reducing it up to 30%. So if we predicted you'll make $1,000, you can see the predictions for a revenue as low as $700, $800, $900, etc, all the way up to $1,300.

We do want to stress out, however, that the predictions are more accurate for the baseline value (the $1,000 in the example). This is because of the natural growth trend your store has, and forcing the increase or decrease in revenue is affected by the lag you have in advertising (the amount of time it takes users to buy since they see your ads). We do the prediction from today forward, and you want to affect sales from today forward... But those sales are also affected by what you did in the last few days (or weeks). Even though we do capture a bit of this behavior looking at your customer journeys, the number of sessions you had in the last period, and estimate the number of people we can expect to come back, this behavior is not 100% captured by this (maybe your refraction period is longer, maybe you had a massive campaign that is still producing results, etc). These simulations are really good, but the best one is the one for the baseline revenue.

Now, we don't estimate only your total revenue, but we actually divide that revenue in two:

1) Returning revenue. This is the revenue that existing customers will bring to your store. Depending on your store, CLV and repurchase rate, this revenue may be very large or very small compared to the total revenue. This is the 'easy money', because customers know you and understand the value of your products, and can be obtained not only by repurchases, but also

2) New revenue. This, on the other hand, is revenue that will be generated solely by new customers. This is harder money to make, but comes with the reward of growing your customer database, allowing you to scale... And make much more money.

Ok, moving on. Once this is done, we pass to next part: the total number of sessions needed to achieve that revenue

b. Sessions

This part is comprised of 2 components. For both we use the help of the paths users follow before completing a transaction in your site (to know more about this, read our article on customer journeys )

a) First, the number of returning sessions you are naturally going to have because of the efforts you did in the last 30 days. It's important to mark that not all of these people are ready to buy yet, and that their coming back to your page may need the influence of other ads, or simply multiple subsequent visits. There are two key aspects to determine how many sessions will be produced, and from where: first, the Bounce Rate of your page. We assume all the bounces are not coming back. We also remove all the sessions which led to a purchase in that period. It's only the sessions that are left after this that we take into account for return. Only then do we refer to your customer journeys: we see the last contact point you had with every user, to know where they are standing, and see the optimal path to conversion from there. This way, we allocate the returning sessions.

b) The number of new sessions you have to bring. In addition to bringing back the old people, you also need to bring new people to your page. Here we turn to customer journeys once again. This time, we see the optimal conversion paths since the very first contact with a customer until conversion, and estimate the new sessions you would have to bring per source. Now, because of the cross-device problem, this optimal paths may not include all traffic sources that are important to you. Your traffic may come mainly from Facebook, but Facebook's influence in purchases may well be lost in the cross device mess. So this approach would say not to spend on Facebook, even if it's a key source for you... How would we notice? And how do we fix it? Because of the growth we are allowing channels with this optimal-path distribution. We look at the traffic sources we are including for these new sessions, and see how much of the total sessions they represent. Then, we look at your historical data. If this new distribution (possible cross-device biased) is resulting in changes too large between channels, we set a cap on the sessions each source can have, and add the neglected channels to fill in the gap. You should know that a) This will only happen if your store has a massive cross-device derived information bias, and b) this 'fill in the gap' distribution is also estimated from your historical data.

c. CPS

Once we have found the number of sessions you have to bring, we can move on to the budget definition. Having the number of sessions per source, it's easy to calculate that source's budget: we just have to look at the CPS of every source.

Now, how do we do it? A source's CPS varies a lot both over time, and depending on the audiences your use. So, CPS is also estimated. We do a general CPS prediction for each one of the sources, using the information of how the CPS changes both over time, and depending on the number of sessions.

d. Reach

As a final step, we estimate how many people would have to see your ads in order to bring you the number of sessions  we said. This is for you to have a general idea on how plausible your goal is: you con know the size of audiences on Facebook, for example, as well as the average number of searches for a word, so you can have an idea as to if it's an attainable number, or if it's way too high (because of the extra revenue you can run the simulation for). Please note, though, that even if we do take frequency into account, there is no way for us to know the 'mixed frequency' of your ads across channels: all we can do is know the reach for each channel, and divide it bye each channel's frequency. However, it's most probable that a user impacted on one source is also impacted by another source. We can only know the intersection for users who do click on your ads, and only when there is no cross device loss. So the intersection between reaches is much larger, and we have no way of knowing its actual value. This is why, we only take the reach of each source, and then sum it together. We do want to stress out that this sum is an overestimated number, and the one you should pay more attention to is the reach per source, not the total reach, because of the intersections between sources

So in summary, we are giving you an accurate insight not only on how much money you should spend in each source, but also how many sessions you need from each one of your sources, including organic traffic, and how big the audience you target should be (this is the paid sources' reach).

If you want to know more or have any questions or concerns, please drop us an email at hello@sensaimetrics.io

When analyzing our traffic and where it comes from, we are used to letting the tools decide who comes from where, how many traffic sources we have, and which are their names.

If someone lands on our page coming from Facebook, we’ll only see them as one more visit from Facebook, regardless of how he found us: for all we can see (at first glance, at least), the user might have clicked on a paid ad, organically found our page, or got to us through an influencer or friend sharing a link.

If we really want to know which exactly it was, we have to filter visits according to UTMs.  However, there are a couple downsides to this approach:

1. We don’t normally know all our UTMs. We have to go to Facebook, look at the campaigns and check for UTMs. And what if we misspelled it in a couple of ads? Our filter wouldn’t include those results, and unless we checked super carefully every link in every ad, we wouldn't even realize.
2. Filtering is tedious when comparing sources. So you've gone trough you ads' links on Facebook and know exactly which UTM you want to look at, and filter the metrics by it. And yes, now you can see exactly the metrics that you want. But once you’ve filtered the results, you've lost sight on the general performance of the page, and of every other source. Influences Jenny brought you 300 visits, but how does that compare to the rest of the sources? How many visits did influences John bring? Was it 150? How about any other type of campaign? You would have to remember all the numbers from all the sources to actually have a good perspective on how the filtered info is doing.

That’s exactly why we built a tool that lets you choose how to divide your sources, so you can see every single feature of Sensai based on such classification.

How does it work?

We read all the UTMs from your traffic sources, and pre-classify them the way any other tool does. However, once we’ve done this classification, there are a couple things that could go wrong: maybe you put in your UTMs ‘mailing facebook’ because it was a mailing campaign created from a Facebook form… Or maybe, it was the Facebook campaign that contained the mailing form. Other than that, you may want to divide each influencer on their own, or maybe you want to group a few together.

So, after our first classification is done, we show you how we’ve divided everything. You decide if you want to move a couple of campaigns, and where to put them. You can create new traffic sources (like ‘influencer Jenny’) and see that source in every metric and every graph in Sensai. Or you can group together things we’ve left apart (like ‘bing’ and ‘yahoo search’), and see them as one source through the entire software.

How do I do it?

You'll see a board of columns with cards in them. Every column represents a source, and the cards in the column are the UTMs we are attributing to that particular source. Also, on top of each column, there is a dropdown menu: this stands for that source’s channel.

You can move cards from one source to another, create new sources and even change the channel each source is associated to. To make it easier, you can see your source UTMs for a more general perspective, or see the Medium UTMs, for a more granular sight and classification.

Seeing all your UTMs this way, it's much easier for you to spot if one UTM was misspelled, and which traffic sources are being misclassified.

How should I divide my sources?

Yes everything sounds great, but how should I divide my sources? Which classification would be more insightful? Which sources should I group together?Which others should I break apart? It depends much on your business and the type of campaigns you're doing with each source, but here are some basic examples to get you started:

Influencers: Influencer marketing is a big part on how stores these days acquire new visitors and customers. Their drive on leads is ever-increasing, so each day more and more shops turn to influencers to promote their products. But as any marketing channel, it needs close watch and a lot of optimization. There are influencers who will work great for you, and one post will be enough for an impressive increase in sales; but there will be others who won’t bring you even one product sale, and who will only end up costing you money.

With traditional tools, it’s very complicated to measure their success, because even if each influencer has their own UTM, the tool will place all those visits/conversions in referral, and there’s nothing you can do about it: you can’t set ‘Influencer Jane’ as a traffic source, and you can’t compare Jane’s traffic to that from Facebook ads, nor analyze it like you would any other marketing channel.

If you are doing influencer marketing, we suggest to create an ‘Influencers’ source, moving all your influencers’ UTMs to that source. This way, you’ll be able to:

• Measure the general metrics and compare their performance to other sources
• See the journeys visitors take after coming from an influencer source. Do they come back doing an organic search? Or do they need another kind of advertising before being convinced to buy? This way, you can know the true cost of an influences-related purchase: if visitors need to come back lots of times, all through paid advertising, maybe influencers don’t work that good for you, or maybe you should consider another kind of content for them
• Measure the assisted conversions they drive
• Know which products/product categories work best with influencers, so for the following campaigns with them you can have a better-chosen product selection.

Also, if you have one big influencer and many small ones, we also recommend to separate him/her from the others. The idea is always for you to get an accurate perspective of what's going on, and if influences John brings you lots of visits and conversions, the overall results will be driven up by him, and maybe hide the poor results of the small influencers... Or vice versa: it's also quite probable that the small influencers are doing good, and as you pay them much less, the overall results are positive; however, even if big influencer John is bringing you lots of visits, that doesn't mean his audience is your audience, and people won't necessarily end up buying.

You should always know who is driving up sales individually, and avoid the blinding effects some sources will have over others when grouped together.

Facebook paid: As we briefly mentioned before, there are many different reasons why someone could come to your page from Facebook. Broadly, we can divide all the visits into two completely different channels (which traditional tools squeeze up together):

• Organic: Visitors who come to your site from organic posts from your Facebook page, or from a link a friend shared, or even an influencer shared.
• Paid campaigns: Visitors who come from paid campaigns you specifically created on the Facebook ads platform, and for which you are paying money, creating ads, selecting audiences, optimizing budgets, etc.

We encourage you to create one source for your paid campaigns and to move your paid campaigns’ UTMs to that source, and let the rest of the traffic in a Facebook organic source. This way you’ll be able to distinguish Facebook’s performance as a purely paid traffic source, and not bring in the noise from people who found your page by themselves, for whom you are not paying money, and who were most likely either already customers, or were already looking for something you had to offer.

Google Ads: Google is one of the paid sources with the most varied types of ads, which can also target people in very different stages of a purchase. When someone is actively looking for your product, brand, or competitors’ brand, people are decisively closer to a buying point, than when they are just browsing youtube looking for funny videos. Even Search and Shopping campaigns vary a lot, and seeing all of them in the same ‘Google Paid’ category is very misleading: you don’t get to realize if shopping campaigns work for your or not, or if the display ads are better. Seeing everything packed up together flattens out the curves, and keeps you from realizing if you should spend a little bit more in retargeting ads and cut down spend on video ads. We recommend a separation between Search, Shopping, Display and Video ads.

Mailing: Usually, you’ll have 2 types of mailing campaigns: the automated ones, where you program a trigger, and when users take a specific action, a very specific email is sent to them. The other type are ‘general’ or ‘newsletter’ campaigns, where you send the same campaign to an entire mailing list, without needing them to take a specific action. These are 2 extremely different kinds of campaigns, that reach users at very different stages, and that should be measured separately. We recommend to divide them in two different sources, so you can understand exactly how visitors from each type of campaign are behaving, and even reduce unsubscription rate by optimizing the time and number of campaigns sent. If you see all campaigns together, and have a spike in visits, you have no visibility whatsoever on which type of campaign is driving up those visits. The same goes for bounce rate. Say the automated campaigns have a huge BR, and the newsletter ones a small BR. If you see them together, you'll probably see a normal BR and move on. However, if you see them separately, you'll realize the automated campaigns have a huge BR. So you can go check your triggers and content and adjust them. This would not only help your BR and unsubrscription rate (we all know if you're bouncing too many emails from a sender, you'll end up unsubscribing)... But also your senders' reputation. This is crucial, because all mailing tools keep a track on your rep. If you go below a certain level, you'll be marked as spam, and service providers like Mailchimp, SendGrid and Sendinblue will stop delivering your emails, and most likely block your account.

In this section, you choose one of the adsets you are running, to see every audience it contains, and how its metrics compare to their best-ever value.

The metrics you see are CPC, CPM and revenue, because those are the most influential for ROAS. As you can see in the example, we are looking at the adset 'int fb ig - Copia'. This adset has many audiences: 'interaccion fb', 'Lookalike (MX,1%) - interaccion fb', 'Interaccion ig', 'Lookalike (MX,1%) - Interaccion ig'... Now, seeing the audiences like this, you can compare the CPC you are having right now with the best CPC value 'interaccion fb' has ever had, and with that of all the other audiences.

The ‘top’ information corresponds to the adset containing audience X, and that had the best ROAS of all adsets containing audience X.

This dashboard helps you understand if a specific audience is suitable for a certain adset: if the current performance across metrics is close to the audience’s top performance, there’s a really good chance that the audience is suitable for the adset, and most likely it will generate a good ROAS.

There will be times when you see the current adset performing better than the top adset. This doesn’t mean it’s wrong: this happens because an audience could have had the best ROAS when the CPC was at .30. Maybe your current adset has a CPC of .20, but the ROAS is much lower. This could indicate many things: for example that this audience is more suitable to instagram, where the CPC is more expensive.

It’s not enough to have a great product, and it’s not even enough to have a great product AND make great content for it. If you’re not targeting the right audience, you might as well be selling ugly, cheap stuff. Say you sell luxury watches of big-name brands, and you pay the best designer in town to make your creatives. If your ads are seen by 16 year old girls, the odds of even one of them buying are virtually 0.

Finding the right audience can be challenging for a number of reasons: the many different kinds of audiences there are, the overlap between audiences, the frequency with which a user from a certain audience should see your ad before he buys, the saturation of audiences, etc. To make things more complicated,  when you go to your Facebook dashboard, the most you can do is see metrics at the adset level. However, normally, each adset has many audiences, so this doesn’t really tell you how each audience is behaving.

Let's take one step back. There are 3 types of audiences:

1. Custom audiences. These audiences can be created based on many things: an email list you have from subscribers (that FB will match to users profiles and serve them ads), from the Facebook pixel based on users’ behavior in your site (like people who visited X page, or clicked Y button, or added to cart); or even from people's interests ('rock music', 'diving') or behaviors they have ('recently married', 'diving instructor')
2. Lookalike audience. These audiences are based on custom audiences. Facebook creates audiences of profiles similar to the ones the base-audience contains. You just choose how similar they have to be, from a range of 1% to 20%  similarity. The idea is: ‘of all Facebook users, find me the 3% most similar profiles to the ones of the people in my custom audience’. The smaller the number, the most similar profiles will be, and the smaller the audience.
3. Behavior/interest audience. For this audience, you choose an interest (like ‘people interested in travel’) or a behavior (‘frequent travelers’), and create an audience based on one or more of those behaviors/interests. You don’t need data from the Facebook pixel, nor a mailing list, which is why this is the go-to type of audience when you don’t have lots of data from your business, when you are opening up a new product, or simply when you are feeling daring/have a little bit of budget left and want to explore new things.

All these different types, as they come from very different places, will naturally have different behaviours. As for audience frequency and saturation, depending on the audiences and products you have, there are different frequencies that work. It’s more likely that for a big or expensive product, people will have to look at you ads several times  before convincing themselves to buy, but for very sensational gadgets, only two or even one glance is enough. Just like when targeting a campaign to people who have already given you their email, they’ll need a different message and different number of interactions than people who have no idea who you are and what you sell.

We built this tool to give you a better visibility on how every one of your audiences has performed in the past, for you to know what to expect in the future, so you can make adjustments accordingly.

Among other things, you can see the 3 best and worst performing adsets for each audience, along with different metrics on how they performed, and the differences between audiences targeted to a specific gender or device.

Say you see ‘Audience A’ which has an outstanding performance for the top 3 adsets, but a terrible performance for the 3 bottom adsets. You see the audience has better numbers for women on mobile, while questionable costs and response from males or in desktop.

You do a bit more digging and check the 3 top adsets. Turns out all 3 had discounts or some kind of promotion. With this whole picture, you now know that this specific audience should be used to target women, on mobile, and only when you have an active promotion.

In the example at the top, you can see that the third audience, ‘ALL_90’, was used for 3 different demographic sections: 2 different age groups, and the possible gender separations. It’s easy to see that when targeted to females, the results are much better - 4 times better! Of course, there are other things you should look at before driving conclusions, like how much money you have spent on each one. If you spent 10x more in the female targeting, then it’s not actually 4x better.

Please note that all the data in this section is based only on data from the past. This section is based purely on data mining, and no predictions are made. It can be possible for an audience that performed very poorly in the past, to show excellent results by changing the demographics or the ad content, as well as it is possible for an audience which has performed great to burn out.

CLV

As you may know (or not, in which case please read this article), CLV is one of the most important metrics for your store. Not only does it help you know your store’s actual value, but it is also key as to how much money you can spend in acquiring customers, for it tells you how much money they’re gonna ‘give you back’.

Ironically, though, CLV is one of the most under-estimated and overlooked metrics. Many startups don’t even calculate it, and many who do only sum all the revenue they’ve ever had and divide it equally between all users.

This approach however not only is pretty basic, but it deliberately lets out a key piece of truth: not all customers are equal. You’ll have people who buy $1,000 from you, but you’ll also have customers who only came in attracted by a massive discount, and won’t buy again unless they again see an insane price drop.

Does it really matter? Absolutely.

First of all, bad customers can be very bad business: by giving huge discounts, and adding all the marketing and operations costs, you can easily be left without profit, or even lose money acquiring those cheap customers.

So, you’ll want to identify the kind of customers who will bring you more money in the long run, and bring more of them, as many as you can. As for the low quality customers… You’ll want to bring as few of them as you can.

However, to do this, you can’t be just looking backwards and seeing how much people end up spending on your site when there’s nothing more to do. To actually take action, you’ll need to know both how much a customer has spent and how much he will spend in the future.

Sounds impossible, doesn’t it? It’s not. We created a tool that tells you exactly how much revenue each one of your customers will bring you in the future, and when they will do it. So you can pamper those good customers, create stronger bonds with them and upsell… And stop worrying and losing money over the bad customers.

The outcome

You can expect 2 very different things from our model:

1) Forward-looking (predicted) values
We predict for you:

• your customers’ average CLV (how much money an average customer will bring to you during their lifetime)
• every customer’s LTV.
• a lists of the top places for you. We’ll tell you exactly how much money you will receive from each demographic location (cities, states and countries)
• list of top 1st products/categories. We’ll find the customers who are expected to spend more money during their entire lifetime. We’ll then find which was the first product they ever bought with you, and which category that product belongs to. This is really big in identifying potential good customers: we’re finding relations between lifetime spend and the first product purchased, so you’ll be able to tell with some degree of certainty, from the very first order someone places with you, if they will be good or bad customers.

2) Backward-looking (historical) data.
All of the above are things we’ll predict from the future… But we’ll also give you insights on what has already happened in your store:

• Monthly new vs old customers
• The places where you’ve sold the most (by country, city and state)
• The products and categories you’ve sold the most.
• Insights on which places, products and categories are rising (selling more than before)  and which are falling (selling less than before)

The actual model

As it does involve quite some math, we’ll explain this in 2 leves of complexity, so if you don’t really want to know exactly all the details of our calculations, you can read only the basic explanation and you’re good to go.

a) Basic

The short version is we do mathematical modeling with your data. Modeling (in math) means watching the past behavior and using it to predict the future behavior of a person or group of people. That’s what we do: we take your past data, feed it to a model, and out come the predictions. We use 2 models: the first one to know when people are going to buy, and another one to know how much money they will spend in each of those orders. The models we use are like superstar for calculating CLV: they were developed by mathematicians and backed up by more than 20 years of research, so you can be assured they work damn well.

b) Advanced

If you want to know a little bit more, we’ll tell you not all modeling is good. If you do it wrong, you can end up with predictions that have nothing to do with reality. For example: if your predictions are way lower than the actual facts, you’ll end up with a massive shortage of inventory, which means you’ll ‘lose’ the money you could had made by selling those units. This is called opportunity cost. If on the contrary you predict sky rocketing sales, you’ll both have lots of inventory that won’t get sold, and most likely also spend lots of money in advertising, so you’ll actually be losing money on two fronts.

Worried? Don’t be. As we said before, the models we use were developed and perfected by PhD mathematicians form universities like MIT, and are featured in many scientific publications in respected science magazines like Elsevier. In this section we’ll explain the first model, called Beta-Geometric model (for short BG). This is the one used to know when and how many times people will place an order. The second model we used is called Gamma-Gamma, but we won’t go over it in this section.

Every single model makes assumptions about the data. These are the cornerstones for defining the whole model. Depending on how accurate (or not) these assumptions are, things can go really well (or not).Our assumptions involve probability distributions. This is where things start to get a bit technical, so if you think you’ve read enough, just skip to the next section.

Before we move on, we want to state a simple but crucial fact for all CLV calculations: not all customers in your database are worth the same, and not all of them are active. There are many who have already bought what they needed, and won’t ever buy again. And it’s ok. It’s only natural. What you want to do, however, is a) stretch the time during which they’ll come and buy (their lifetime), and b) make them buy as much as you can during that period (their lifetime value, or CLV). That’s the notation we’ll use: active are the customers who will still make another purchase with you. Churned, lost or inactive customers are the ones who won’t ever buy again.

Ok, so, moving on. These are 5 key assumptions we make:

1. Customer activity. We assume that, for every customer in your store the behavior is: first the buy a little. Then they see your product is good, or see your store is not a scam, or they’ve already seen your catalogue looking for the initial item and found another one they liked, so they’ll buy more. It’s only natural, they first try it, and when they know it, or even know of it, they get confident and spend more. After giving you lots money, though, their activity starts declining, until they eventually stop buying from you. This behavior is seen in every single e-commerce. You may think yours is different… It’s not. What IS different is how fast customers will go from spending a little to spending big to going back to spending little. Another thing that does change is how big those leaps are: how much more they spend when they spend more, and how much less they spend when they spend less.
   All of this is modeled by what is called ‘Poisson distribution’. Basically, what we say is that for every customer, his purchases will be in the shape of a curve like the one of the following:
   

Which curve, though, that will depend both from that particular customer, and from your business model. Poisson distribution has 2 parameters, λ and k, which determine what we could call the ‘amplitude’ and ‘wavelength’ of the curve.

Here, λ is the transaction rate of your customers, ie, how many times they buy in a given period of time. When you model something using a probabilistic distribution, what you do is called ‘fitting the parameters’ λ and k, to find the pair that best adapts to every customer.

1. Transaction rate similarity. We assume every single customer behaves differently. He/she is unique and so are their habits. However, there is some pattern of behavior: you’ll probably have seen that you have many customers who don’t buy much, and a small chunk of customers who spend lots and lots of money with you. This heterogeneity also follows a probability distribution, the Gamma distribution.
   

This curve is similar to the Poisson one. This distribution also has two parameters, and it will model how the λfrom the Poisson distribution varies across customers. A way to understand this is that the horizontal axis is how many times they buy from you (eg: 5 purchases), and the vertical one how many customers fall into that category (how many of your customers place 5 orders during their lifetime).

1. Inactivity. We assume the inactivity of a customer comes right after the last transaction they made. If someone bought from you son May 1st and never came back, we’ll say he became inactive May 1st after his purchase. Now, how can we know in advance when a customers’ last purchase will be? Once again, we use probability, and once again a probabilistic distribution enters the scene. Now it’s a geometric distribution:
   

The geometric distribution is a special case of the Poisson. Here what we basically assume is that after the first transaction, probability of churning (or never returning) is much higher tan after a couple (or a lot of) purchases. The probability is always greater than 0, but is much smaller as time goes by. Again, this is a fair assumption for the more a customer buys with you, the more loyal he becomes to the brand. After all, there’s a reason why (s)he keeps coming back, right?

1. Inactivity similarity. Again, even assuming every customer is different, because of your business model there will be some similarities between you customers. For example, if you sell mattresses, you’ll expect customers to come back and buy much less than you would if you were a food delivery service. Mathematically, this means that the probability of churning also follows a probabilistic distribution. This time, it’s a beta distribution
2. Customer uniqueness. Every customer is absolutely unique. If Mary suddenly decides she doesn’t want any business with you, that does not affect Bob’s purchases. Likewise, if you convince Tony that your brand is absolutely fantastic and get him to spend lots of money, that won’t affect John’s purchasing behavior. So, the transaction rates and the probabilities of churning vary independently between customers

If you want to know more about what you can do with your CLV and how to make the most of it, read this article. If you have any more doubts either regarding the models, what you can do with them, or any other subject for that matter, please feel free to contact us!