The Unicorn Builders’ Guide to Business Health Metrics

Sami Yabroudi, Founder, Principal at Flatiron Data & AI

August 2025

We often develop metrics and frameworks in new industries where there isn’t an accepted best practice. Even within established industries we find ourselves addressing crucial blind spots. Here is our “first principles” guide.

This guide focuses on the top-most level of business health analysis. Drill-downs can be found in subsequent guides.

Start by Asking The Questions Out Loud

Internalize your true priorities by asking these out loud:

• Robustness of offerings — “Why are customers buying from us? Will this continue? Why do we lose customers? How do we strengthen customer loyalty and spend?”

• Consistency of outcomes — “Are we delivering operationally where we need to? What are our key product deficiencies?”

• Trending financial soundness — “Is our gross income positive? What is the payback on marketing dollars spent? When will we reach full profitability?”

• Effectiveness of internal culture and talent — “How effective is our organization compared to corporate headcount?”

We strongly recommend asking these questions out loud in conversation; simply reading them doesn’t have the same effect

New customer growth is usually the last item on our list — marketing dollars can easily masks a brittle business barely held together.

Before Metrics: List Units of Analysis, Plan Out Personas and Other Key Splits

Units of analysis — each suited to different questions — include transaction, customer, and cohort. Top ways to segment — i.e. split — analyses are by customer persona, geographic market, and offering type+tier. These segmentations can expose much larger and insightful effect sizes than an overall average and strongly inform strategy.

Customer Persona is Usually the Most Critical Segmentation

A company’s customers rarely all have the same needs, constraints, and responses. MarkStrat alums will remember the D2C Sonite and Voditearchetypes, for example. A B2B vendor can segment clients by size and industry, among many other variables.

Persona development may be split further into marketing personas and customer personas. Once a marketing lead becomes a customer, a company typically has significantly more information to more accurately classify the customer.

Persona development and refinement is a continuous process. At the “quick and dirty stage,” qualitative hypotheses made by revenue team leaders — perhaps also incorporating user interviews — may be effective enough. A data team can quickly test classifying customers into those groups and see if the groups distinguish themselves in behavior. We always advise incorporating as much external data as possible (ex: census income averages tied to ZIP code) to increase the accuracy of categorization.

Clustering is an algorithmic approach to developing personas. In addition to intrinsic attributes, customer behavior sequences can also be inputs to these models. These models require no human-generated hypotheses.

Retention is the Most Encompassing Metric

Retention measures whether customers are returning, and usually represents the single best quantification of competitive value being offered. In some cases referral rate may substitute.

Time-based cohorting provides a strong start for understanding D2C retention. A north-star company OKR might be “Retention of a cohort after 2 years is 55%.”

Dig Deep ASAP into the Key Value(s) to Customers

This inquiry is critical, and yet often kicked down the road. Without it retention will simply be a black box. There is absolutely no reason for a company to guess — perhaps incorrectly or/and incompletely — at the value that current and potential customers are looking for. Misalignment is extremely costly and may not be rectified until it’s too late (we’ve seen some things…).

Customer interviews, surveys, and focus groups kick off a strong value delivery framework. In these interviews, avoid leading questions and neutralize bias in analysis (firstly by diversifying the analysis team). A data team’s role in this process is to help in design level inquiries, characterize effect size of reported phenomena, and estimate whether the majority of customers have likely been represented in the qualitative research conducted.

Measure the Key Values Offered, by Hook or by Crook

Once competitive values have been determined, delivery of these values must be measured in every manner possible. One might have to get creative, tool additional information capture within a software platform, scrape the internet, or accept imperfect proxies.

For example, if a ticketing app’s users are mostly looking for the best price, continuously scraping competitors’ prices and understanding pricing position in real time is crucial. If, instead, the app’s users are compelled by rich event discovery, begin by measuring the frequency of various behavior paths and their outcomes.

Quantify Loyalty and Switching Costs; Monitor Conditions for Customer Departure

Continuously quantify the potential value of evolving competitors to current customers, as well as the likelihood of the market reaching dangerous switchover conditions or step increases in CAC.

These models can also incorporate approximations of stickiness and loyalty via cross-sell, upsell, free resources, incentives, deep integration, etc.

Accurately Monitoring “Churn Reason” is of Highest Importance

In D2C and SMB B2B industries it can be tempting to treat churn as natural and not dig in with the care of an archaeologist. However, we would advise investing in a combination of operationalized churn reason surveys, churned users <> key events analysis, and periodic user interviews (to confirm that the former 2 are not structurally biased or out of date). Letting customers dictate company strategy can be significantly cheaper and more effective than alternatives.

Retention is a Lagging Indicator; Stay Ahead with Leading Indicators

It may take a few months for an unexpected drop in retention to become clear; for starters, a month-based cohort model can’t show anything within the most recent 30 days. Obviously, an earlier signal would enable lucrative preventative corrections.

Counting churns on a given day definitely provides an earlier signal. But oftentimes an even earlier signal can be found, such as a subscription pause, a bad service experience, a focused search that did not lead to a purchase, or any other event that is found to correlate with a churn down the line.

This feature article, “The Unicorn Builder’s Guide to Business Health Metrics,” is by Sami Yabroudi and was originally published by Flatiron Data & AI. It is republished here with the author's and publisher's permission. To dive into the full piece, visit the Flatiron Data & AI blog.