Five Ways the Square Lifecycle Team Uses Cohorts

For today's How I Amplitude, we're joined by Anas Farah, Lifecycle Data Scientist from Square. At Square, with over 70 products in their product suite, cross-sell opportunities are an incredible growth lever for the team.

But like any context-based offer, timing and audience is everything.

In this How I Amplitude, Anas walks us through five real-world examples that showcase how Square’s lifecycle team’s transition to self-service analytics has supercharged their cross-sell targeting.

His team works closely with the Product Marketing team to identify which audiences to test specific offers with, starting from when they only have qualitative data all the way through to how they feed their ML models and predictive cohorts.

Our mandate: Self-service PMMs

I joined Square back in 2022 with a clear mandate: Help the growing team of 100+ product marketing managers become self-sufficient with audience segmentation.

The goal

Our PMM team works to guide sellers to all the various products in Square's suite that could fill their particular needs—from the flagship point-of-sale product to appointment booking, payroll management, banking, and more. Getting the right customer the right offer at the right time means responsive segmentation.

The state (and how that made life hard)

At the time, PMMs depended heavily on data scientists to create audience segments—a slow, resource-intensive process that wasn’t ideal for anyone involved. And it certainly got in the way of rapid testing and building feedback loops.

The solution

I spent about a year working closely with PMMs on their audience creation needs and evangelizing Amplitude cohorts. I educated them on our martech stack, event dictionary, and instrumentation processes in order to help bridge the gap between them and engineering teams.

Did it work?

Short answer: Yes. Square’s PMMs have progressed from data team dependents to confident, self-sufficient power users who build out audience cohorts for various marketing campaigns with ease.

Long answer: Yes AND the examples here show off our self-service approach in action, ranging from simple announcement campaigns to sophisticated behavioral targeting that combines ML with PMM expertise.

Case 1: Sending a quick announcement

When we rolled out the Square Card in the UK, our PMMs wanted to message all active sellers in Great Britain to announce the launch.

Before Amplitude, general announcement blasts required data science support to write SQL queries that would sync audiences through our internal reverse ETL tool. While these requests required minimal data team effort, they were typically low priority and could take one to two weeks to complete.

With Amplitude, PMMs were able to build these audiences independently in minutes by defining an audience cohort as:

• basic user properties (country_code = GB) and
• Square's standard active seller criteria:
  • at least one completed payment in the last 91 days and
  • no frozen or deactivated status

We synced the cohort directly to our campaign platform, and set it live!

While simple, this self-service capability reduced lag time for announcements and helped get the PMMs acclimated with using Amplitude cohorts independently. More importantly, the PMM team had a lot more freedom to run more tests and pick up their own velocity of experimentation.

Case 2: Using context clues to cross-sell

A big part of our work on the lifecycle team at Square is cross-selling new products and services to our existing customers.

While our ML models typically provide the best targeting results for these efforts, new product launches don’t usually have enough data to build effective models. In these cases, PMMs turn to two alternative approaches:

• Contextual cross-sell: Using information we already have about a seller (like size, location, or type of business) to determine if a product would be a good fit
• Behavioral cross-sell: Looking for specific actions or patterns in how sellers use existing products that might indicate they'd benefit from another product

A straightforward example of a contextual cross-sell came when we rolled out Square Shifts, a suite of workforce management tools for sellers. To build early-release hype, PMMs wanted to send a contextual cross-sell to sellers who had ten or more employees

The challenge was that we didn’t have any information logged on a seller’s employee count in any of our user fields. However, by exploring contextual data in Amplitude, the PMM team found a proxy measure: an optional onboarding question that asks sellers about how many employees they have. This was close enough to the criteria they’d wanted, and because it existed in Amplitude, they could also segment by it easily. Just like that, they had an audience cohort to contact.

Additionally, they set up the campaign to go out 45 days after a seller initially joined this audience to give them time to fully onboard with their first product before getting the cross-sell for this new one.

Case 3: Optimizing cross-sell timing with behavioral data

PMMs typically progress through a journey of increasing sophistication with Amplitude:

1. First, they build basic announcement cohorts
2. Then, they experiment with contextual targeting using demographic data
3. Finally, they graduate to creating behavioral cohorts based on usage signals and patterns

This last stage—building behavioral cohorts—has become the most common and valuable use of Amplitude for our PMMs.

Take, for example, our launch of Square Checking, a checking account service for sellers in the US.

Our Banking data science team found that sellers who attempted three or more Instant Transfers within a 30-day period and had processed at least $10,000 in total payments over the last year showed strong propensity for adopting Square Checking.

The team also discovered that this behavior was most strongly associated with sellers in our services category, who showed the highest likelihood to adopt Square Checking after having used Square Instant Deposit.

Using Amplitude's cohort builder, our PMMs could now create a highly targeted audience of sellers who:

• Successfully completed at least 3 Instant Transfers in the last 30 days and
• Were based in the US market and
• Belonged to the services audience category and
• Had processed $10,000 or more in payments over the last 365 days and
• Met Square’s standard active seller criteria:
  • Had card payment activity within the last 91 days and
  • Weren't currently frozen or deactivated

Case 4: Combining ML models with PMM expertise—a hybrid approach

Our PMMs often want to layer their own targeting logic on top of our ML models.

Take our Instant Deposit cross-sell campaign. Our ML model had identified sellers likely to adopt Instant Deposit, but the Instant Deposit PMM had learned through research that sellers who adopted Square Card first tended not to need Instant Deposit as much, so he wanted to add that logic to the campaign.

Using Amplitude Cohorts, the PMM was able to build an audience that combined:

1. The ML model's prediction for Instant Deposit adoption and
2. Their own rule to exclude sellers who had interacted with Square Card and
3. Standard criteria like country restrictions and active seller status

By combining the broad pattern recognition of our ML models with the specific product knowledge of our PMMs, we were able to run a successful adoption campaign that didn’t waste resources on an audience subsegment we knew wouldn’t be fruitful.

The hybrid approach isn’t just good for results—it’s also been a huge win for building PMM’s trust in our ML models. ML can efficiently identify large quantities of likely adopters, but PMM was understandably skeptical of the models beating the quality of their manual targeting. By letting PMM layer their own targeting criteria on top, it’s easier for them to see the strengths of ML, and they now feel like the hybrid approach gets them the best of both worlds.

Before letting PMMs add their own criteria on top of ML predictions, we do validate that it won't hurt performance. But if we don't see any degradation in performance, we're happy to let them refine the targeting this way. It's another example of how we try to empower our PMMs without undercutting our own ML models.

The key technical piece that makes this possible is how we structure our ML scores in Amplitude. To make it simple for PMMs, we don't show them the raw scores—instead, we use a dynamic threshold that changes daily. Any seller that crosses that threshold gets marked with a "1" to indicate they're likely to adopt the product.

Then we sync this value into Amplitude as user property which PMMs can easily use. There's no need for them to understand the complexity of the ML model—they just see it as another toggle they can use in their audience building.

Case 5: Bridging public web data with lifecycle marketing

Audience cohorts also help our PMMs run their own experiments.

The Square Appointments team wanted to A/B test offering a hardware discount of 20% as a way to get new users to sign up for Appointments. Half of the people who visited the Appointments landing page would see the discount offer while the other half wouldn’t, and later receive the discount code via email if they signed up.

The challenge here was instrumenting an event that would combine our Optimizely experiment data (which would split the audience 50/50) with our public web tracking for signups. Getting all this to work together would have required significant engineering time that we just didn't have.

But we found a creative solution using Amplitude. All of our Optimizely data already flows into our Amplitude instance, so we built a cohort that did the following:

1. First, we identified everyone who viewed this experiment (using experiment ID and variation ID from Optimizely)
2. Then we filtered for people who specifically created an account through the Appointments variant within 14 days of seeing the experiment
3. Finally, this cohort would receive an email with a code to redeem their 20% hardware discount

What's valuable about this approach is that we didn't need to create any new custom events or wait for engineering resources. We were able to leverage existing event data in a new way to enable this experimental approach. This is a great example of how we often work with our public web teams to bridge their experiments with our lifecycle marketing efforts.

Putting this into practice

We’ve been able to do incredibly sophisticated work with Amplitude at Square, but our work has also proved that you don’t need ML models or data teams to do effective research on audience segmentation. For smaller teams, you can use the same audience segmentation principles from my first three examples:

1. Start with basic targeting, like active users in a specific region
2. Layer on contextual information, like business type or size
3. Add behavioral signals, like frequency of certain actions

Using Cohorts, you can test different combinations of these criteria to find what works best for your product.

If you want to dive into self-serve analytics, my recommendation is to start simple with campaigns you're already running.

Instead of sending that next audience request to your data team, try building it yourself. Then gradually add more sophisticated criteria as you gain confidence. Look at your power users and what makes them unique: their common behaviors, how frequently they take certain actions, or which features they use together.

Everything is an experiment, and now you have the tools to run them yourself.

Bonus section: What's inside Square's lifecycle machine?

Our team manages all in-app, email, and push messaging experiences that sellers receive from the moment they join Square.

Besides our flagship point-of-sale (POS) product, we now have an entire suite of products for our sellers including appointment booking, payroll management, and banking.

Our global lifecycle marketing program goal is to guide our sellers through the various additional products that may benefit them at their current lifecycle stage.

At the heart of our operation is our internal customer data platform, which ingests data from multiple sources, including:

• Square's iOS and Android apps
• Squareup.com website activity
• Logged-in dashboard experiences
• Third-party tools like Amplitude, Optimizely, and Iterable
• Our Snowflake data warehouse

A significant portion of this data flows into Amplitude, where it is organized across different projects and combined into cross-views used for cohort analysis, creation and syncing. Rather than relying on a third-party CDP, Square built an internal CDP to manage identity resolution, ensure data privacy, and maintain full control over event processing at scale.

Q: Why did Square build their own CDP instead of using a third-party solution?

We had specific needs around reliability and scale—processing billions of events daily requires precise control. While our CDP is inspired by Segment's architecture, we needed customization around privacy, identity resolution, and specific connectors not available in third-party tools. The build decision came down to three main factors:

• Control over connections and infrastructure
• Meeting our specific service-level agreement (SLA) requirements
• Cost efficiency at our scale of events

That said, third-party tools often have better governance features, which remains a gap in our custom solution. It's a tradeoff we continue to evaluate.

Q: How does Square handle B2B marketing at scale?

Our self-serve motion is technically B2B since we're marketing to businesses, but given we message millions of sellers annually, we often use B2C-style approaches in how we schedule and structure communications. We segment based on Gross Processing Volume (GPV)—at certain thresholds, sellers move to a separate, sales-led motion using different tools like Marketo. While this creates some complexity, it allows us to appropriately serve both small businesses and larger enterprises. We're working on improving synergies between these two motions.