When I was asked to help out at a Series B SaaS company to build their marketing function, I quickly realized they had a pretty unique challenge. The business had grown almost entirely by word-of-mouth — no structured marketing, no sales playbook. But after riding that wave of organic growth, things had stalled, and the investors were getting anxious. I was brought in to figure out how to turn that early success into something scalable.
The first step was to identify who our best customers actually were. We needed to get specific about which customers were worth focusing on, so we developed a model around three key concepts: Play, Stay, Pay.
It’s like trying to find the golden apples in an orchard — you have to understand which trees bear the sweetest fruit, not just the ones that look good at first glance. So, we created a composite score based on these three metrics and then sliced it across company size, vertical (industry), geography, champion seniority, department, and title.
To build out these customer profiles, we needed to enrich our data with more detailed information. We tried several providers — ZoomInfo, Lusha, PeopleDataLabs — but Apollo turned out to be the best in terms of match rates. However, we ran into an issue with industry classifications. The default categories from the enrichment vendors were too broad. “Information Technology” could mean anything, from SaaS companies to IT consulting firms.
So we got creative. We built our own verticals and then used AI to reclassify companies, tweaking and iterating until we had more meaningful categories. As one colleague said during this process, “It’s like trying to organize a garage — every time you think you’ve nailed it, you find a whole new pile of stuff in the corner.” But in the end, we finally got a clear view of who we were really serving.
Our analysis revealed that the sweet spot was mid-market companies, with VP-level champions in six specific industries. These customers had the highest composite score when you factored in ACV, churn rates, and conversion rates to paid. This formula allowed us to zoom in on the right opportunities, quickly.
What’s interesting is that once we had this data, it became obvious. Suddenly, these clusters jumped out when we looked at new customers or when we thought about our longest-standing, most enthusiastic users. It was like seeing a pattern emerge where you hadn’t noticed it before — almost like watching the dots connect themselves.
Of course, as soon as we presented this data, some folks on the leadership team pushed back. They argued that our data might be misleading. One executive said, “The companies in your dataset may have started out as small businesses and grown into mid-market firms — so you’re misclassifying them now.”
To test this, we sliced the data by the timeframe in which customers had signed up — specifically looking at those who joined in the last 12 and 6 months. While the changes were minor, our core conclusions still held: mid-market companies with VP-level champions in key industries were consistently our most valuable customers.
Identifying who our best customers were was a great first step, but we needed to figure out why they were so valuable if we were going to craft compelling messaging for them. So, we tackled this from two different angles: human-driven exploration and LLM-powered insights.
We gathered 24 colleagues from various departments — sales, customer success, data science, marketing — and split into teams. Each team focused on a different ICP cluster, tasked with answering several questions about their customers, such as:
This was a little like detective work — each team pieced together clues about what drove our customers’ decisions. One outcome from this process was surprising: we realized how hard it was to get a full picture of our customers because so much of our data was scattered. As one participant noted, “It’s like we have half of the pieces of the puzzle, but we don’t know what the final picture is supposed to look like.”
While the human-driven exploration was useful, it had its limits. We didn’t have comprehensive data for every customer. That’s where LLMs came in. We had recordings of sales conversations from about 10–20% of our paying customers, so we used that as a foundation. Here’s the process we followed:
The consistency and depth of insights were impressive. It was as if the LLM had taken a magnifying glass to the data and revealed patterns that were harder to see at first glance.
The real magic happened when we combined the human-driven insights with the outputs from the LLM. The teams provided context, a nuanced understanding of the business, while the LLM offered consistency and scale. Together, we created a stronger, more actionable ICP and refined messaging that felt more precise and targeted.
The outputs from the ICP project became the backbone of our product marketing efforts, feeding into Message House and campaign briefs. By leveraging LLMs to speed up the process, we were able to quickly move from identifying ICPs to campaign execution, significantly shortening the time it usually takes to go from concept to launch.
As a colleague mentioned, “It’s like taking a rocket ship when everyone else is still walking — once you have these tools in place, everything moves faster.”
Leveraging LLMs for core product marketing workflows has transformed how we operate. By combining rich, full-bandwidth data with detailed prompts, we’ve accelerated our ability to identify who our best customers are, why they matter, and how to speak to them. This process doesn’t just speed things up — it improves the depth and quality of insights, setting us up for better, more effective campaigns in the future.
As we continue to capture more unstructured data, I’m confident we’ll find even more ways to create workflows like this one. And that’s what makes the future of product marketing so exciting — getting smarter and faster every step of the way.