The project was called Fusion, a new B2B payment feature at PingPong, the cross-border payments company where I worked from 2021 to 2025, as Marketing Manager, Ads and Lifecycle at the time. The feature was designed to grow through existing customers. Our import–export customers were already paying their suppliers through the platform, and the bet was that those suppliers, who first met us when a payment landed in their account, could become customers themselves. I ran the North America side of the launch, which meant the record design, the routing, and the lifecycle emails. We opened on a targeted pilot, somewhere between forty and eighty accounts filtered out of a few hundred existing customers on three Salesforce fields, and it ran a quarter or two. I worked with product in China, sales leadership in the U.S., and every record passing through three systems: Salesforce, HubSpot, and SalesLoft. Everyone reviewed the launch plan. The risk sat in the data.
The week the reports stopped agreeing
A referral program routes people into your funnel through paths your systems were never designed to recognize. A referred supplier's first touch is a payment email from someone they already do business with. If your systems can't say who they are and where they came from, every downstream report is fiction. And every team involved is working from a different fiction.
I know because my first attempt produced exactly that. Early on, I let referred suppliers flow into the standard lead process, and the numbers stopped agreeing almost immediately. In HubSpot, marketing's view counted them alongside organic signups. In Salesforce, the same people sat in the pipeline as stalled leads, because a supplier who showed up to receive one payment doesn't behave like a lead, and the pipeline velocity looked wrong to everyone reading it. Two reports in the same week told contradictory stories about the same humans, and both were “correct” by their own definitions. I'd seen it before. Blend the data once and nobody untangles it later. So I stopped scaling and fixed the record first.
A dedicated environment for referred users
So I stopped treating referred users as leads. I gave them their own source and environment: a distinct lead-source value in Salesforce, matching tags in HubSpot, and, the piece that mattered most, a lookup field linking every referred supplier back to the customer whose payment introduced them. Referral entry points carried UTMs that resolved to the referring customer's account, and a Zapier step I built wrote the referrer onto each new record. That step ran for over a year and went to marketing ops when I left.
That one field paid off immediately. My automations pre-qualified referred suppliers and dropped them into the right SalesLoft cadence, on the same Zapier layer I built out to 200+ workflows. Handoff dropped from roughly 12 hours of manual routing to one or two minutes, read off Salesforce lead-created and owner-assigned timestamps across about a month of records. I could finally see which customers were driving adoption, and referral volume clustered hard around a small set of them, roughly ten accounts carrying most of it, the ones whose transaction volumes and partner networks made referring realistic. So I handed sales that shortlist along with the decks and one-pagers, and they ran the conversations. Every record traced to its origin, so I could measure what the business cared about: did this supplier complete a real payment after KYC. The conversion figures stay internal (fintech has opinions about that). The shape I can give you is order of magnitude: low hundreds of referred suppliers entered, and sales worked those accounts.
Marketing reports are often built on raw data pulled from multiple systems, stitched together with naming conventions that only make sense to the person who created them.
How you name and tag records tells the people who weren't in the room what the report means. The person reading the report next quarter (finance, support, a new hire) shouldn't need the report's author present to interpret it.
Ship the MVP
The other discipline was scope. The launch stack I actually shipped was four assets:
- Retargeting audiences on Google Ads and LinkedIn, referred visitors who didn't convert stayed reachable, run inside the always-on paid budget with no line of its own
- A dedicated landing page, one destination, one measurable entry point, instrumented with GA4 and Hotjar
- A basic one-pager, enough for customers to explain the offer for us
- A SalesLoft cadence, written to read like a person checking in
Everything else waited for evidence. I judged the landing page on behavior before form fills, watching scroll depth and exit patterns in GA4 and Hotjar, because with a new product, pushing traffic into an unready sales flow creates noise for everyone. And because I'd kept sourcing clean from day one, I could see what the MVP was and wasn't doing, and I would have added pieces when the behavior data called for them. It never did. Those four assets were the whole stack.
Write the state machine first
Referred suppliers didn't follow the standard lead lifecycle. Pretending they did was what broke the reports in week one. So I wrote the actual states down. A referred supplier enters as a limited user: usage caps on otherwise open features, so they could experience fast transfers and the FX dashboard before committing to anything. They become a full customer only after completing KYC. And they count as a success only after a real payment moves on the platform. I wrote a cadence for each state: nurture for leads, activation for limited users, onboarding for new customers, each with a distinct goal. Not every message pushed for action; some I wrote only to maintain momentum.
Sales deliberately entered late. A few key accounts had told their account managers they were uncomfortable with us contacting their suppliers too early, so the early stages showed suppliers the product through retargeting and opt-in email, until one registered on their own. At every stage, the trigger did more work than the copy. Lifecycle systems work best when you define the state and the trigger more precisely than the words.
On Fusion I defined the record before anyone briefed a designer or wrote a subject line. Everything else sat on top of it, including the report both teams finally trusted.