How Meliz drives privacy-first growth through Airbridge MMM

Introduction

Meliz is a one-stop fashion search platform that provides easy access to various fashion items from best selling brands in the world.

Overview

meliz, a fashion meta-search platform based in Korea, used Airbridge’s proprietary Marketing Mix Modeling (MMM) solution to analyze marketing performance across various channels, learning that reallocating its marketing budget would increase app installs by 5%.

A single search for all fashion brands

meliz is a one-stop fashion search platform, providing easy access to various fashion items from globally renowned brands.

It aims to become the ‘Top of mind’ shopping portal for all Korean fashion shoppers, with the ultimate goal of becoming a global gateway to fashion shopping.

Operated by ABLY Black, a subsidiary of a leading fashion commerce platform in Korea, meliz now offers over 5 million products from 7,000+ brands on its search engine. The service has shown an incredible growth rate — it has reached 800K+ MAUs (Monthly Active Users) and 1.6M+ cumulative app downloads within a year since the official launch.

Challenges

Examining the true contribution of marketing channels while protecting user privacy

After making great investments to acquire new users, meliz now sought to maximize its growth potential.

Yet, as the user privacy protection measures, such as Apple’s App Tracking Transparency (ATT) policy, have been widely adopted, the existing last-touch attribution could no longer accurately depict the marketing channels’ contribution as in the past.

Thus, meliz needed a new solution to successfully understand each channel’s credit while securing user privacy to prepare itself for the post-privacy era.

Solutions

An intuitive MMM solution for agile decision-making

For the past few months, Airbridge has closely worked with meliz to conceive a viable MMM model that can successfully measure channel-level marketing performance for the post-privacy era.

During the project with meliz, Airbridge offered a proprietary frequentist MMM model which continuously upgrades itself using day-to-day datasets, optimizing its analyses and predictions. With the aggregated daily spending on UA (User Acquisition) campaigns for each channel as the main input, Airbridge measured the contribution of each channel to the number of app installs and recommended a budget plan to maximize the budget efficiency. The analysis was run on the 6 months of historical data, from January 2022 to June 2022.

Also, since the MMM model is inherently built to detect the correlation, not the causation, it requires a regular calibration with the causal inference results (e.g. real-world experiment) to enhance the model performance. Yet, since meliz did not have historical experiment data for calibration, Airbridge offered to use Incrementality results, which have proven to resemble the real-world lift experiment, as an alternative. The calibration was conducted for the last week of May.

👉 Please refer to the attached link for further explanation of Airbridge’s Incrementality solution.

Results

Results

Meliz gained some crucial marketing insights to better understand each channel's true contribution and plan the future budget to increase efficiency.

• 33% less organic contribution compared to last-touch attribution
• 39% more installs credited to DA(Display Ads) channels compared to last-touch attribution
• 5% potential increase in installs with the optimized budget plan suggested by the MMM model
• 4% potential to decrease the average CPI with the optimized budget plan suggested by the MMM model

* Organic refers to the contribution of non-marketing activities.

Organic contribution

Since the rollout of iOS 14.5 and the implementation of ATT, the mobile app industry has faced difficulty in attributing the opt-out ATT user events. meliz was not an exception, meaning that it has seen a great increase in ‘unattributed’ installs - the install events that are not attributed to any of the marketing activities like the organic - based on last-touch attribution.

By nature, the last-touch model is unable to attribute the opt-out user events unless the user changes the consent status. In other words, unless the user opts in, the user event remains unattributed.

Unfortunately, MMM also has its limitations. As explained earlier, the MMM model trains on the correlation between marketing activities and marketing performance. This poses challenges to measuring organic results, or the contribution of non-marketing activities. The silver lining is that the MMM model can be calibrated with external data like experiment results, adjusting channel-wise contribution to reflect reality. Airbridge used Incrementality results, which resemble real-world experiments, for calibration in this project, and thus, generated a much organic contribution than last-touch attribution.

This shows that unlike the last-touch model, which is more reliant on user-level data, the MMM model can successfully measure channel-level contribution by incorporating various external experimental data even in the post-privacy era.

Increase in the contribution of DA channels

Last-touch tends to overestimate the contribution of search ads that involve lower funnel activity - the search, while underestimating the display ads with higher funnel activity - the exposure. Unlike the last-touch attribution, MMM focuses on the correlation between each channel and the target events (install), and thus, can fairly assign credits to each channel. In this study, Airbridge MMM gives 39% more credit to display ad channels that have usually been neglected under last-touch attribution.

Potential to increase budget efficiency

Since the MMM is a time series model, it can forecast and simulate the scenarios in the near future. During the analysis, Airbridge’s MMM recommended a new budget plan to invest more in cost-efficient channels to expect a potential increase of 5% of total installs with the same budget. Also, based on Airbridge’s recommendation, the average CPI is also expected to decrease by 4%.

Other than the key business insights, the Airbridge MMM model itself has also proven to be a viable and reliable solution as below:

• 2 weeks of onboarding required to deliver the initial results since data ingestion
• 3 months of historical data required as a minimum to set up an initial model
• 7% average prediction MAPE during the analysis period (from March to June 2022)

* MAPE (Mean Absolute Percentage Error) indicates the model’s prediction error in percentage terms, with lower MAPE meaning better model accuracy.

Airbridge MMM Features

With its proprietary machine learning engine, Airbridge's MMM model analyzes the relationship between different marketing elements. The model trains with the aggregated channel data while incorporating a variety of data processing methods, such as the carryover effect, to better reflect reality.

With its model, Airbridge offers two MMM reports — “Marketing Mix Analysis” and “Budget Optimization” — to help clients make data-driven decisions. The Marketing Mix Analysis report analyzes the contribution of each advertising channel to desired analysis targets such as revenue or installs. “Budget Optimization” proposes an optimal budget allocation to achieve maximum performance under given budget constraints.

Marketers can freely test various marketing scenarios using these features on a user-friendly interface. Furthermore, Airbridge’s integration system and own API endpoint support marketers to easily ingest the data for analysis and export the results from the reports to facilitate the practical use of the MMM solution for day-to-day decision-making.