Success story

MebelVia decreased the Cost of Sale by 11.5% using AI-driven multi-touch attribution

Learn how a large online furniture retailer decreased the Cost of Sale by 11.5% using multi-touch AI-driven attribution.

Request a demo
MebelVia decreased the Cost of Sale by 11.5% using AI-driven multi-touch attribution
Company name: MebelVia
Industry: E-commerce
Location: Eastern Europe

“Investments in SegmentStream paid off in just 1 month”

— MebelVia

The company

MebelVia is one of the largest furniture online retailers in Eastern Europe with millions of website visitors per month.

It acquires a significant part of its customer base through paid advertising channels such as Google Ads, Facebook, Criteo, and others. At this scale, it is extremely important to analyse the performance of paid ads and constantly test new optimisation approaches to decrease the Cost of Sale.

In the process of doing so, the MebelVia team faced three main challenges that needed to be solved to achieve the maximum performance of their online advertising.


1. Attribution problem

MebelVia customers interact with many marketing channels and visit the website multiple times before they finally decide to buy.

Yet, single-channel attribution models such as Last Non-Direct Click attribution are not able to fairly distribute the value of conversion between multiple channels. Especially when it comes to offline conversions such as sales via phone calls.

As a result, MebelVia has overinvested in some channels and underinvested in others.

2. Quality of data in marketing reports

With millions of website visitors per month, MebelVia started experiencing data sampling issues in Google Analytics. Also, the previous marketing report automation tool used by MebelVia was unstable and led to data discrepancies and unreliable results.

On such a large scale, minor details can have a significant impact on overall data quality. MebelVia wanted to ensure that they could trust the data in their marketing reports.

3. Audience segmentation in retargeting

MebelVia heavily relies on retargeting campaigns to drive users back to the website. But when millions of people visit the website each month retargeting becomes quite expensive. It is necessary to understand which users have a high or low probability to buy so that retargeting costs are optimised and maximum performance from the spend achieved.

Prior to using SegmentStream platform, MebelVia manually segmented all website visitors into different audience groups depending on their website behaviour. However, the segmentation process took too much time while the performance wasn’t good enough.

After an in-depth evaluation of different vendors, MebelVia and their digital agency Profitator decided to partner with SegmentStream to overcome all of these challenges.


  • Automate marketing reporting while ensuring the highest quality of data. Data in the reports should be always up-to-date and automatically updated daily. It should not be sampled nor have any other limitations;
  • Implement a fair multi-touch attribution model to measure the true ROAS of each marketing channel and quickly evaluate new marketing campaigns without incurring high costs;
  • Automatically understand which website users have a high or low probability to buy and use this data to optimize retargeting campaigns across multiple channels.


Overall solution architecture Overall solution architecture

Step 1: Automated marketing data collection into Google BigQuery

MebelVia has lots of marketing data sources:

  • 7 different advertising platforms (Google Ads, Facebook, Criteo, Yandex, etc.)
  • website where customers can place an online order.
  • Google Analytics as the primary web analytics tool.
  • Calltouch as a call tracking tool to match incoming phone calls with website sessions.
  • RetailCRM to store and manage all orders, both online and offline.
  • Google Sheets to store cost data from marketing sources that do not have an API (SEO, Email, etc.).

To be able to build a unified marketing reporting all online and offline data should be collected into one place first. MebelVia decided to choose Google BigQuery as its marketing data warehouse.

Thus, as the first step of the solution, all data has to be automatically imported into MebelVia’s Google BigQuery account on a daily basis and without any data loss.

1.1. Cost data import automation

First, SegmentStream automated the import of all the cost data from multiple advertising accounts into Google BigQuery:

  • Google Ads
  • Facebook
  • Criteo
  • Yandex.Direct
  • Yandex Market
  • MyTarget

The setup process took around 15 minutes as SegmentStream provides turn-key integrations with the most popular marketing data sources.

Cost data from all advertising sources are automatically aggregated in a unified UTM-grouped format. MebelVia doesn’t have to do any transformations manually to get the data in a proper format.

Further, it is not necessary to mark the campaigns with specific UTM parameters to stitch campaign data with sessions on the website. SegmentStream can parse the UTM mark-up that the client is using.

SegmentStream can also import cost data from marketing sources that do not have an API (Organic Search, Email, etc.) via Google Sheets.

Table with costs data Table with costs data

2.2. Website user behaviour collection

SegmentStream does not rely on Google Analytics to collect website behaviour data.

Instead, SegmentStream has its own JavaScript tracker to collect website user behaviour data. This has the following advantages:

  • There is no limit on the amount of data collected (for example, Google Analytics has a limit of 500 hits per session).
  • The website behavioural data can be collected in real-time.
  • There is no limit on custom dimensions and metrics.

As soon as the SegmentStream tracker was implemented on the MebelVia website, real-time data and events started to flow directly to BigQuery.

Fragment from a raw data table in BigQuery Fragment from a raw data table in BigQuery

3.3 Offline data import

Orders received from incoming phone calls are the main offline conversions for the MebelVia business. Calltouch is used for dynamic call-tracking, while RetailCRM is used to store and manage orders.

SegmentStream has turn-key integration both with Calltouch and RetailCRM.

These integrations made it possible to automatically collect this data in Google BigQuery and match each order to specific website sessions, including orders made by phone.

Step 2: Automated data transformation

After all the data has been collected, SegmentStream automates all necessary data transformation needed to aggregate and prepare the data for the visualisation and reporting.

The overall data transformation process for MebelVia consists of 11 automated daily imports from its advertising systems and offline sources into Google BigQuery and 12 SQL queries.

To ensure that each transformation starts in the proper order every day, the SegmentStream platform has built-in automated and fault-tolerant data transformation processes (ETL Workflow Management System).

Examples of data transformations

Once a day, raw user behaviour data is transformed into sessions. This helps to save money when building reports in BigQuery.

The hit table contains data for one day (around a gigabyte). BigQuery pricing depends on the amount of data processed in the cloud. Complex hit table queries process higher amounts of data, increasing the cost of using BigQuery. Queries become significantly cheaper If the necessary data is aggregated into a session table in advance.

Comparison of the cost of raw and aggregated queries in Google BigQuery Comparison of the cost of raw and aggregated queries in Google BigQuery

Advertising costs are matched with website sessions using a combination of 6 parameters: date and 5 UTM tags.

User identifiers (userId, Google clientId, SegmentStream anonymousId, phone number, etc.) and the time of the conversion are used to match offline conversions with website sessions.

Step 3. Model training for AI-driven multi-touch attribution

Performance metrics of the first version of the machine learning model Performance metrics of the first version of the machine learning model

There is more information about the AI-driven multi-touch attribution model in this article on CXL.

SegmentStream analyses user behaviour on the website in real-time and predicts the buying probability of each visitor. Predictions are based on large amounts of historical data of user behaviour and purchases.

To train the machine learning model, we use various behavioural signals about all user interactions with the website. These include:

  • product image views
  • clicks on various functional elements
  • use of filters
  • search queries within the site
  • adding to cart
  • clicks on phone numbers
  • phone calls

The model also uses information about the frequency of the events, the time they occurred, and the numerical characteristics of some micro-conversions. Using all this data the model learns to predict purchases in the future.

The first version of the model was trained a month after the start of data collection. At that point, it was possible to build AI-driven multi-touch attribution. The model is retrained as new data is collected, and the accuracy of predictions is improved.

Step 4. Preparation of unified marketing reports based on AI-driven multi-touch attribution

Change in purchase probability after each customer visit Change in purchase probability after each customer visit

Using a mathematical model, SegmentStream measures the buying probability for each user at both the beginning of the web session and at the end of the session. The difference shows an uplift in the conversion probability and therefore the real contribution of this traffic source.

The value of the conversion is now fairly distributed among all traffic sources based on their real contribution to this conversion. As a result, Mebelevia now gets a comprehensive marketing report in Google Data Studio with Cost of Sale data based on AI-driven multi-touch attribution.

Marketers also want to know how the Cost of Sale (CoS) will vary depending on the attribution model. SegmentStream tables were originally built to compare multiple attributions in a single report.

Example report with demo data: Source/Medium + Campaign analysis based on SegmentStream AI-driven multi-touch attribution compared to Last Non-Direct Click attribution model. Example report with demo data: Source/Medium + Campaign analysis based on SegmentStream AI-driven multi-touch attribution compared to Last Non-Direct Click attribution model.

Step 5. Bid management automation based on AI-driven attribution metrics

MebelVia uses the K50 bid management system to automate bidding strategies across Google Ads, Yandex.Direct, and other channels.

Previously, K50 used the Last Non-Direct Click attribution model for bid optimisation. After the SegmentStream implementation, MebelVia and its digital agency Profitator switched to AI-driven multi-touch attribution.

This was possible as SegmentStream is also integrated with K50 and can automatically export AI-driven attribution metrics into the K50 system for further bid optimisation.

Step 6: AI-driven audience segmentation

The SegmentStream machine learning model can predict conversions in real-time in user browsers. These predictions are used to automatically segment users into different groups by their buy probability in the future.

The real-time model divides users into 3 groups:

  • Low probability of purchase (Down)
  • High probability of purchase (Up)
  • Very high probability of purchase (Up Up Up)

A new website user automatically falls into the “Down” segment. While using the website, the visitor performs various actions such as viewing the lists of products, using filters, sorting products in the lists, searching and so on. After each visitor’s action in the browser, the model recalculates the probability of purchase in real-time.

When the probability of a purchase exceeds a certain level, the visitor enters either the “Up” or “Up Up Up” segment. Information about the visitor segment is then sent to Google Ads, Yandex.Direct and other marketing tools.

All of this means that marketers can adjust their bidding strategy according to the user’s buy probability. For example, reducing bids for the audience not interested in buying, or increasing bids to attract users with a high probability of conversion.

This dramatically helps to increase the efficiency of retargeting campaigns. Marketers no longer need to spend the same amount of money for users with different conversion probabilities.

Also, these predictive audiences can not only be used in retargeting campaigns but can be activated in look-alike campaigns as well. For example, marketers can target new users that are similar to existing users with a high-probability to buy.

As the next stage, together with MebelVia and their digital agency Profitator, we plan to run A/B-test in Google Ads and Yandex Direct to see the real impact of such bid adjustments on the final Cost of Sale and ROAS metrics.


MebelVia started the bid optimisation process based on AI-driven multi-touch attribution on January 23, 2020. The first results appeared in February.

Figures for February 2020 and 2019 were compared since there were no changes in the channel structure in 2020 compared to 2019. The result was that ad spend in February 2020 had decreased by 11.5% year-over-year without any loss in revenue.

This level of cost saving has meant that the annual investment by MebelVia in the SegmentStream platform was covered in just 1 month.

CoS decreased by 11.5% compared to February 2019 with no decrease in revenue. CoS decreased by 11.5% compared to February 2019 with no decrease in revenue.

Customer testimonial

I have been responsible for running MebelVia online marketing for two years and during this time we have faced various problems named above. The biggest one was the attribution.

At some point, we started to face inefficiency in our marketing mix using Last Non-Direct Click attribution, and more and more questions arose. We had been thinking of various approaches for optimising advertising campaigns for a while.

Looking for options, we found several solutions on the market that provided customized multi-touch attribution models. Typically, such models are manually configured by specialists, which, in my opinion, makes them less trustworthy. As soon as something changes on the website, the whole model has to be redone manually, and this is a real issue.

At this time, the SegmentStream team suggested trying a new approach — AI-driven multi-touch attribution, which is fully customized for our business, yet it is 100% automated. It sounded pretty exciting as it eliminates the “human factor” problem.

We trained the model, and together with the SegmentStream team, we integrated AI-driven attribution with our bidding management system K50. For the past few months, we have been planning and allocating our budget according to the AI-driven multi-touch attribution.

We are delighted with the results! Thanks to the SegmentStream team for a high-quality product and professional work.”

— Yaroslav Semenov, Profitator Digital Agency (Kokoc Group)

“Very often I hear that all digital agencies are only interested in commissions. It was really nice to find out that this is not true. Working closely with Profitator, we realised what a real agency is: fighting for every increase in ROAS for the client and ready to try innovative technologies to achieve maximum performance.”

— Constantine Yurevich, CEO, SegmentStream

You might also be interested in

More success stories

Get started with SegmentStream

Request a personalized demo with our team!