Traditional consumer segmentation doesn't reveal the true customer; it merely replaces one massive, inaccurate mean value for the entire population with six or seven smaller, equally flawed sets of mean values. By forcing individuals into rigid cohorts, we indulge in a form of crass over-summarisation.
A customer trapped inside a segment doesn’t actually resemble the "average" profile of that segment at all, it usually means nothing more than that they happen to sit closer to that specific mathematical centroid (the behavioural centre-point for the given group) than any other.
In reality, the difference between being placed in Segment A versus Segment B is often incredibly marginal. It simply isn't wise to build an entire content strategy on what for some customers is the algorithmic equivalent of a coin flip.
The Descriptive Fallacy: Characterisation vs. Definition
The danger is amplified because marketers routinely misunderstand what a segment profile actually represents. They treat segment descriptions as if it were a strict definition, when in reality, it is merely a characterisation.
When a persona deck states that a segment is "The Young Urban Professional," it is describing the mathematical average of that cluster's attributes. It is not a gatekeeper rule.
Whilst a segment will over-index for the characteristics in the description, there will still be a lot of people in there who don't match many (or perhaps even any) of the criteria.
For example, a holiday company using this segment might target the segment with city-breaks to nightlife hotspots, assuming they are all 25-year-old renters with no children. But this is likely to be a poor match for the 55-year-old homeowner who, despite their age and location, happens to be in that segment because their media habits and spending patterns are mathematically closer to that centroid.
An AI model, by contrast, would ignore the segment caricature and use the individual's actual behaviour, such as frequent online searches for rural retreats, to correctly identify them for a different, more relevant offer.
When a marketer looks at the catchy title of a segment and assumes everyone inside it fits that exact mould, they are falling for an illusion. They are targeting a caricature, not a customer.
From Micromanagement to Governance: Letting Go of the Map
To be fair, segmentation became an industry standard for valid historical reasons. In an era of limited computing power and sparse data, it provided a simple, explainable framework that aligned with the capabilities of older marketing technology. It gave marketers a sense of control and a tangible way to answer the vague but essential requirement to "understand their customers."
The problem is that this approach forces us to ask the wrong question. Instead of "who is our customer?", we should be asking, "for what purpose do we want to understand them?"
For high-level proposition design, a broad consumer segmentation is perfectly adequate. But when you are communicating with existing customers - for whom you hold rich data on past transactions, intentions, and predicted outcomes - it is clearly insufficient.
This demands a shift in mindset: from control to governance. Marketers need to let go, not of their strategic goals, but of the need to manually execute every decision. It’s like the old request to "map out the customer journey."
In an advanced marketing function, the potential paths are orders of magnitude too complex to ever draw a complete map. Instead of trying to micro-manage every communication, the modern marketer's role is to establish the principles, rules, and governance that allow automated systems to operate intelligently within safe boundaries.
Lessons from Credit Risk: The Move to Combined Models
Interestingly, this realisation isn’t just bubbling up in marketing. My colleague Nick Sime recently explored a parallel paradigm shift in the world of credit risk modelling in his article, "Combined vs bespoke models in credit risk: Does segmentation still add value?"
For years, risk modelers relied heavily on segmentation - splitting populations into separate buckets like "new vs. existing" or "prime vs. sub-prime" - to build discrete, localised linear models.
However, Nick highlights that with the advent of Modern Machine Learning (ML), such as Gradient Boosting Machines (GBMs) and Deep Neural Networks (DNNs), traditional segmentation actually hinders performance.
Nick's core conclusion is profound: when advanced models are trained on a single, combined population, they naturally outperform separate, segmented models. Why?
Because advanced ML models are exceptionally good at capturing non-linear relationships across the entire dataset without needing humans to draw arbitrary lines in the sand.
In Nick's words, in the machine learning era, segmentation is unlikely to improve results and will often make them worse. A single combined model is consistently stronger and vastly simpler to govern.
Putting Segmentation in its True Place
If the highly regulated, mathematically stringent world of credit risk is abandoning segmentation, why is marketing still lagging behind?
It is time to put segmentation in its rightful place. We need to stop using arbitrary marketing segments to decide which specific offers a customer receives. In the age of real-time AI agents, continuing to rely on static buckets is inexcusable.
We should be using the actual, granular behavioural records we hold for each customer to drive individual, automated decisioning, using tools like Recurrent Neural Networks (RNNs) for next-best-action or collaborative filtering engines for recommendations.
This is not to say that segmentation is useless. It still has an important, though strictly bounded, place:
- Strategic Marketing Planning (Consumer-Based Segments):
Broad, macro-segmentations are useful for high-level strategy, helping leadership understand market landscapes and guide long-term proposition design. This is a top-down exercise in intent, not a tactical execution tool. - Health Monitoring & Evaluation (Behavioural Segments):
High-level behavioural groupings - such as Recency, Frequency, Monetary Value (RFM) frameworks - are effective macro-level dashboards for monitoring the health of a customer base and evaluating major initiatives. - Paid Media & Acquisition (The Only Execution Exception):
We must allow for segments within paid media (e.g., lookalike audiences on Meta or Google). Prospecting is a consumer-based activity where we don't own the data, forcing us to rely on the platform’s aggregated cohorts to bridge the gap.
Beyond the Centroid
The moment a consumer becomes a customer, the reliance on segments must stop.
Just as Nick pointed out that a single combined model yields optimal performance in risk, the same applies to marketing. Feeding actual, un-summarised customer data into deep learning engines removes the artificial friction of segments. It allows the AI to discover hyper-nuanced relationships a human looking at a "Persona Profile" would never dream of.
Segmentation still has its place, but whilst using it for execution was a good interim solution for an era when data was sparse and systems less capable. But today, using six centroids to actively manage the experiences of six million people isn't "customer centricity" or "personalisation"; it’s a legacy of old constraints.
It's time to embrace a new role, using and governing the systems that let modern AI treat our customers like the unique individuals they are.