Welcome back to another lively episode of Funnel Lab Fridays! If you’re just tuning in, Funnel Lab Fridays is your go-to LinkedIn Live series where we delve into the challenges and innovations that data-savvy marketers face in today’s fast-paced world. I’m your host, Doug Bell, CMO at CaliberMind. Joining me today are two rock stars in the field: Jordan Crawford, the mastermind behind Blueprint GTM, and our very own Nic Zangre, CaliberMind’s VP of Solutions Architecture and a self-proclaimed RevOps Jedi Master.
Today, we’re diving deep into a topic that’s near and dear to every marketer’s heart: How to Use Data to Validate Your Ideal Customer Profile (ICP). We’ll explore the contrasting methodologies between using Large Language Models (LLMs) and Machine Learning (ML) to fine-tune your ICP, and how these approaches can revolutionize your go-to-market strategies.
Setting the Stage: The Battle of Methodologies
Doug kicks things off by introducing the session’s theme. He sets the stage for a friendly showdown between two distinct approaches to validating your ICP: one championed by Jordan using LLMs, and the other by Nic leveraging traditional ML.
“We’ve got sort of opposite methodologies using forms of AI to get the same result,” Doug explains. He assures us that despite the tech-speak, the conversation will be anything but dry.
Meet the Panelists
Jordan Crawford is the Founder and CEO of Blueprint GTM, a company that helps businesses define their ICP programmatically. Jordan’s approach involves using LLMs to scour public data, structure it, and qualify companies based on real-time information. This method not only identifies who to target but also provides insights into why they’re a good fit.
Jordan humorously acknowledges his return to the show: “Thanks for having me back. I don’t know what poor decisions were made internally to have me back, but I’m so grateful for them.”
On the other side, we have Nic Zangre, a seasoned veteran at CaliberMind who recently transitioned into a new role. Nic elaborates: “I’m stepping into a VP of Solutions Architecture role to help navigate all the snafus, help them see around corners, and really help them adopt and implement attribution and marketing analytics.”
The Fuzzball Challenge: Turning Abstract ICP into Actionable Data
Doug recounts how the CaliberMind team, after extensive customer interviews and data analysis, developed a well-defined ICP. The challenge? Translating that “fuzzball” of an ICP into actionable strategies.
“We handed you an ICP with the following boundaries,” Doug tells Nic. The ICP was B2B companies that are sophisticated—meaning data-savvy or marketing-sophisticated—and heavily dependent on marketing for growth.
Nic accepts the challenge: “Machine learning models are limited to the training data. So, Doug and the team tasked me with building this model and some of the signals they wanted to run on, we didn’t have readily available, so we had to make proxies for these signals.”
Nic’s Machine Learning Approach
Nic dives into his methodology:
- Hypothesis Formation: Identifying proxies for hard-to-measure signals like marketing sophistication. For instance, the number of web pixels and JavaScript tools running on a company’s website could indicate their marketing tech stack’s complexity.
- Data Gathering: Leveraging Clearbit data within Salesforce to enrich account profiles. This data includes technology tags, employee size, industry, and more.
- Data Flattening: Converting complex data into machine-readable true/false signals. “Machine learning likes ones and zeros,” Nic explains.
- Model Building: Using Google BigQuery’s ML capabilities to create a logistic regression model. Nic highlights how BigQuery simplifies what used to be thousands of lines of Python code.
- Model Validation: Assessing the model’s accuracy and false positives to ensure reliability.
Nic showcases the results, emphasizing the model’s ability to explain the weight and attribution of each signal. “We know that if they have an ABM vendor running on their site, they’re already very sophisticated,” he notes.
Jordan’s LLM Approach
Switching gears, Jordan presents a contrasting methodology using LLMs and tools like Clay and AI agents.
He shares a real-world example involving a client targeting HVAC trade schools:
- Data Collection: Starting with a free list of all cities in the U.S., totaling around 33,000.
- API Utilization: Using Serper API to search for HVAC trade schools in each city. “It’s $0.75 for 1,000 requests. There are a lot of decimals in that number per request,” Jordan quips.
- Agent Deployment: Sending AI agents to find relevant contacts at these schools. He gives the agents specific instructions and context to ensure they retrieve useful data.
- Data Verification: Employing a second AI agent to review the first agent’s work, improving accuracy.
- Cost Efficiency: The entire process costs less than a penny per row, making it highly scalable and affordable.
Jordan emphasizes the importance of starting with the smallest possible segment to ensure relevance. “It’s a lot easier for a human to understand both the question to ask the model and if the model is doing a good job,” he says.
The Pig Metaphor and the Niche Strategy
In a humorous moment, Jordan holds up what appears to be a microphone but is actually a gigantic pig figurine. “This is my pig of a process,” he jokes, using it as a metaphor for his method.
He underscores the effectiveness of targeting niche segments. By focusing on a specific subset—like HVAC trade schools in California—Jordan can ensure high relevance and better engagement.
“I think now a lot less about total addressable market scoring and trying to make sense of the entire market,” he explains. “With my clients, I try to do the other thing, which is, what is the nichiest niche of a niche segment?”
Comparing the Approaches
Doug notes that while both methods aim to validate the ICP, they cater to different scenarios:
- Nic’s ML Approach: Ideal for larger companies with extensive datasets. Machine learning thrives on large amounts of data to train accurate models.
- Jordan’s LLM Approach: Suited for smaller companies or startups that may not have vast amounts of data. By focusing on niche segments, they can quickly test and refine their ICP without the need for complex models.
Nic acknowledges the merits of Jordan’s strategy: “I like Jordan’s methodology on creating these niches because you’re able to prove it out in one and then kind of extrapolate your learnings to others.”
Data Quality and Utility
Both panelists agree that data quality is paramount.
Nic states: “Garbage in, garbage out—not to sound cliché, but machine learning generally thrives well when you have enough training data.”
Doug adds: “Just because the data is there doesn’t mean it has utility. A big part of what we’re doing is getting the data in a state where it has utility.”
The Role of RevOps and Data Science
Nic highlights the emerging intersection of RevOps and data science: “It’s the combination of RevOps and data science—we know it’s coming. So you’re just ahead of the curve in many ways.”
He mentions that CaliberMind is now building a hybrid model, integrating tools like Clay to find more exotic signals not available through traditional data providers.
Final Thoughts and Key Takeaways
Doug wraps up the session by emphasizing that the choice of methodology depends largely on the company’s size and available data:
- Large Companies: Should leverage their extensive datasets using machine learning to fine-tune their ICP.
- Small Companies or Startups: Might benefit more from Jordan’s niche-focused approach using LLMs and AI agents to quickly validate and act on their ICP.
Jordan leaves us with sage advice: “Come in with an opinion. Hire someone like a Doug that can help you make sense of the right questions to ask. If you don’t ask the right questions, the large language models will lie to you with conviction.”
Nic adds: “It’s exciting to see tools like Clay to augment what we’re already doing.”
Closing Remarks
Doug thanks both panelists for their invaluable insights: “Thank you for all your time. Thanks for being with us today.”
As the session concludes, Doug gives a sneak peek into the next episode, featuring Emma Phan from Maven Clinic, discussing Improving Win Rates with AI Summaries for Sellers.
Key Takeaways
- Methodology Matters: Choose your ICP validation approach based on your company’s size and data availability.
- Data Quality is Crucial: Both ML and LLMs require clean, relevant data to produce accurate results.
- Niche Targeting: Focusing on highly specific segments can yield better results, especially for smaller companies.
- Embrace New Tools: Innovative platforms like Clay and AI agents can streamline data collection and validation.
- Ask the Right Questions: Your models are only as good as the questions you pose. Expertise in framing these questions is invaluable.
