‍ZALORA is a giant in the eCommerce world

Part of Global Fashion Group, ZALORA is Southeast Asia’s Leading online fashion retailer. With an inventory of over 3,000 brands, ZALORA serves its 50M+ monthly visitors with a wide range of products, from apparel to shoes and accessories, to home and living products.

ZALORA owes much of its meteoric growth to its open embracing of new technologies to solve its business challenges.

‍

User Reactivation: a known enemy

Keeping active customers engaged is easy. 

To hit their aggressive growth goals, ZALORA knew that they needed to find a way to drive additional purchases from those users who had gone inactive (See: Order frequency has 2x the impact of average order value), and traditional user journeys weren’t going to do it.

With this goal in mind, ZALORA partnered with Aampe.

‍

Messaging “The Aampe Way”

With Aampe, CRM teams don’t spend time building and testing user journeys and trying to figure out “what message to send when.” Instead, they focus on truly understanding their users and writing the most effective and impactful messages possible.

Aampe’s model handles the rote tasks that are better suited for AI/ML, like determining the ideal message timing, frequency, and content for each user, while delivering CRM teams unparalleled insights about their users, which they can use to further refine their messaging across all of their marketing channels.

‍

To build their messaging library, ZALORA used Aampe’s GPT-3 Integration to quickly generate over a hundred thousand unique, instrumented messages for a range of products, from lipstick to sportswear, aimed at different buyer motivations from popularity and affordability to convenience and quality.

After these messages were generated, the ZALORA team let Aampe’s model identify and deploy the messages at the timing and frequency and with the content the model determined had the highest propensity to drive action from each user.

‍

Topline results

While ZALORA has been working with Aampe since November of 2021, let’s zoom in on a 30-day period:

Over this period, ZALORA used Aampe to automatically send and optimize the content, timing, and frequency for nearly 12 million messages to over 1.1 million customers across three different countries.

During this time, Aampe messaging drove:

  • 13.2% incremental additional app visits
  • 12.8% incremental additional add-to-cart events
  • 6.7% incremental additional purchase events

…across those three countries during that same period. To be clear, this was all incremental value — meaning, these were visits, adds, and purchases ZALORA would not have seen if they had not used Aampe.

‍

How does Aampe attribution work?

Aampe employs a control group as the basis of attribution estimates. We measure the success rate (visit rate, add-to-cart rate, purchase rate, etc.) for messages sent through Aampe and then create comparable rates for the control group. We then discount our success rates by the control success rate. So if Aampe messages are associated with 100 purchases, but the control group success rate is half as large as the test success rate, we only attribute 50 of those purchases to Aampe messaging.

You can read more about how we structure our control groups here.

‍

A closer look

To maximize our re-engagement effort, we first needed to understand the types of users we need to focus on, so we split up ZALORA’s users into the following groups:

  • New. The user was first seen on the app within the last seven days.
  • Active. The user last made a purchase within the previous seven days.
  • Recent. The user last made a purchase in the previous 8-14 days.
  • Lurking. The user last completed a purchase, if at all, over 14 days ago but visited the app within the previous 14 days.
  • Inactive. The user last visited the app over 14 days ago.

The three most critical groups that were identified to help increase order frequency were New Users, Recent Users, and Lurking Users. We also identified the Abandoned Cart use case as critical to increasing order frequency (as, since the user has already taken the steps to open the app, find a product, and add it to the cart, the effort should be minimal to push them over the finish line to purchase).

Let’s explore those critical users and use case:

‍

New User Activation

ZALORA knew it was critically important to identify their future power users as quickly as possible, but how could they do this practically?

Aampe works by adapting messages based on individual user responses, but how do we know what to message new users (for whom we don’t have this data)?

Just because a particular message, send time, or frequency may have historically worked well for a substantial group of users doesn’t necessarily mean that it will work well for any particular user, so for the first week of a user’s life on the app, Aampe randomly assigns messaging choices to rapidly learn user preferences.

(From our data, these random assignments produce about the same results as an app’s typically “one message fits all” blast messaging. The difference is we’re actively learning from each of these responses or lack of responses.)

Over users’ first 35 days on the app (The first week of learning plus a subsequent four weeks), Aampe can see several common patterns emerge:

  • Consistent positive preference. After the initial learning period, Aampe can consistently identify and send messages that align with the user’s preferences.
  • Consistent negative preference. After the initial learning period, Aampe can’t identify what users like, but we can identify what they don’t.
  • Inconsistent preference. A user may initially respond positively to the messages we’re sending, but then they may shift and start responding negatively (this often happens when a user is simply not interested in buying on a regular cadence, so after getting the thing they want, all of their preferences turn negative). These inconsistencies can be high-variance — going from a strong positive preference to a strong negative preference or can be low-variance — going from a weak positive preference to a weak negative preference.
  • Dropout. Some users become un-messageable shortly after downloading the app because they turn off notifications. A dropout may be a bounce (e.g., They are messaged the first day and then turn off notifications almost immediately). Most other dropouts happen after the first week of messaging.
  • Flatline. For some users, the model simply never gets a good read on their preferences. Messaging patterns for these users are similar to what we would get from randomly picking messaging choices.

And the following table shows the performance of different new user messages on a representative user group, based on the first-35-day messaging pattern Aampe was able to establish.

‍

Why identifying potential power users early is important

As you can see from the table, when Aampe is able to identify and exploit a user with a consistently positive preference, we were able to drastically outperform as compared to the other new user groups.

Approximately 32% of new users over the course of the 30-day period exhibited some degree of consistent positive preference, and those new users accounted for over 65% of all purchases and 65% of all revenue seen from all new users over the same period.

Outside of the consistent positive preference users, the next most successful group was the Inconsistent, High Variance group. (users for whom Aampe can identify a positive preference at some point add more value than other users, even if that high positive preference doesn't last for a long time.)

How does Aampe personalization work?

Within the bounds specified by ZALORA, Aampe was able to vary message timing (both the day of week and time of day), content (which catalog item they messaged about), and copy (value proposition, call to action, tone of voice, etc.) for each user.

Aampe then learned how different individual users responded to all of these messaging and timing choices — using machine learning to infer responses where there were none — and then stored those preferences as “personalization scores”. 

Personalization scores are the probability that a user will respond with a valuable action (purchases, adds to carts, app visits, etc.) if sent a message that incorporates a particular timing, content, or copy choice. We used the value of these personalization scores to identify the new-user messaging patterns covered above.

You can read more about personalization scores here.

‍

Re-engaging recently-active users

The two other high-value groups that ZALORA identified for our representative group were Recent and Lurking users. Again, during this period, 10% of the messages were sent to Recent users, and 70% of the messages were sent to Lurking users (Reference the messaging column in the table below):

Note: The volume of messages sent to each user group was influenced by multiple factors including our model’s determination of the individual user’s messaging timing and frequency preferences as well as the total number of users in each group — in any app, there are typically many more Lurkers than Active users

*The “Visits”, “Carts,” and “Purchases” columns show the percentage of each event that took place for each group within 24 hours of a user being messaged.

‍

Note that, despite Recent users only receiving 10% of the total messages, they were responsible for an overwhelming 40% of app visits, adds-to-cart, and purchases (We were also able to bring Lurking users back. They also made up ~40% of the purchases over this time window) directly accomplishing ZALORA’s goal of increasing order frequency from groups who had purchase potential, but were not purchasing before being messaged by Aampe.

‍

How did Aampe messaging succeed where traditional user journeys and ad hoc messages failed? 

Click here to learn more.

‍

Solving the abandoned cart issue

Research from Baymard Institute shows that 69.82% of all online shopping carts are abandoned, costing eCommerce stores $18 billion in lost sales revenue every year. The situation at ZALORA was no different.

While ZALORA was sending abandoned cart messages that were effective shortly after the abandonment occurred, they knew there was still a significant opportunity to convert users well beyond the first 24 hours since the cart abandonment event.

The following chart shows the probability that a representative ZALORA user who viewed a specific item on the app without purchasing it will come back and purchase that specific item within seven days of the original view.

‍

Each line tracks the probability of eventual purchase given different messaging influences:

As expected, the graph goes down and to the right (The longer it’s been since a person has viewed an item, the less likely it is that they’ll actually come back and purchase it), but what’s more interesting is what happens before the 0% probability point.

Note: All propensity calculations are derived directly by the Aampe model.

‍

  • The bold, black line towards the bottom is when no notifications are sent. There’s only a dismal 0.35% probability that someone will come back and purchase the item on their own after one days’ time.
  • The lowest grey lines (solid and dashed) represent the standard and generic marketing and lifecycle “blast” notifications that ZALORA was sending through their CRM platform. The marketing messages do slightly better at just below 0.6% effective at 24 hours, but they fall off quickly, and after less than three days’ time, they’re no better at converting users than not sending any messages at all.

‍

The top four lines are Aampe-generated messages:

  • The red lines (solid and dashed) are general messages sent with Aampe messaging and timing optimization, but they don’t contain any specific item recommendations. These generally perform better than the generic “blast” or standard CRM messages, but again, by 3.5 days after the item has been viewed, their effectiveness is relatively indistinguishable from the “no messaging” strategy.
  • When Aampe sent a message pointing to a specific item (even if it wasn’t the exact item the user abandoned) was more effective (starting at a .75% probability and staying consistently more effective than the rest of the strategies all the way out to the seven-day mark)
  • …but what was most effective by far (almost twice as effective as standard lifecycle messaging) was when Aampe sent a message about the specific item that was abandoned. It was still converting at 0.35% a full 3.5 days after the product was viewed!

‍

If specific item messages are so effective, why doesn’t Aampe send them all the time?

Again, it all comes down to user preferences. Some users prefer messages about specific items while others prefer more general or discount messaging. Just because one method is more effective overall, doesn’t mean it’s the right choice for everyone.

Hey, that sounds like A/B testing, doesn’t it? Read here to learn why Aampe is more effective than traditional A/B testing.

‍

Using data to make smarter decisions

For the last step, which makes CRM even more strategic for ZALORA’s organization, Aampe is able to surface insights about the messages that ZALORA is currently sending to this reference group — and not just CTRs, but further downstream data as well. Important metrics like revenue and add-to-cart rates for this experimental group, which can be attributed to specific Aampe messaging.

‍

‍

And ZALORA is able to drill further down into this data to uncover insights about how their specific messaging about their products would directly affect revenue.

For example, from the data below, you can see that these users are responding to messages about lipstick affordability much more strongly than they are about a value proposition about quality — and this insight can be carried forward into the rest of ZALORA’s advertising, website messaging, etc:

‍

‍

Aampe can also give ZALORA insight into their users in general to understand what messages are resonating strongest overall:

‍

Aampe essentially turns ZALORA’s messages into a survey tool that allows ZALORA to actively learn about their customers through the messages they send, without any additional effort.

Increasing effectiveness while reducing workload

All of these results are great, but your ROI will quickly plummet if it takes your team substantial effort to achieve them.

In addition to offloading user journey building and message selection and scheduling from the ZALORA team, Aampe directly tied into ZALORA’s systems to enable ZALORA to send hyper-relevant messages about specific products based on relevant conditions (like proactively messaging users when a product they’d previously viewed dropped into a low stock condition, which is their most successful campaign to date).

For example, with Aampe, a single CRM employee is able to strategically manage the daily selection and sending of millions of messages to millions of users from a messaging catalog of thousands of unique messages — and see and assess results like those above — in only a couple hours a month.

‍

With such powerful tools and insights, it’s no surprise that ZALORA is Southeast Asia’s #1 fashion eCommerce retailer, is it?

If you want to learn how you can get these tools for yourself, drop us a note at hello@aampe.com.