The test that matters
Most operators have segmentation. Much of it is not doing any work.
The test is simple: does this segment cause us to do something different? If two groups receive identical communications, identical offers and identical treatment, the distinction between them is a description rather than a segment.
Applying this test to an existing segmentation scheme usually finds that a substantial proportion of it exists because someone built it, that several segments overlap heavily, and that the ones actually used in targeting are a small subset.
The corollary is that a small number of well-chosen segments beat a comprehensive scheme nobody uses.
Value, and why averages fail
The dimension every operator segments on, and the one most often handled badly.
Gambling revenue is distributed with extreme skew. A small minority of customers generates a large majority of revenue, as established in the iGaming Basics course. In that distribution, the average customer value figure describes nobody.
The practical response is deciles or similar divisions: rank the base by contribution and divide into ten equal groups. What emerges is usually striking. The top decile may generate more than the bottom eight combined. The gap between the top percentile and the rest of the top decile may itself be large.
Several things follow.
The groups behave differently, not merely spend differently. Session patterns, product preferences, channel responsiveness, price sensitivity and tolerance for friction all vary by value tier, which means treatment should too.
Offers should be calibrated to the tier. A bonus meaningful to a customer in the fourth decile is trivial to one in the first and expensive across the population if applied uniformly.
Aggregate reporting on the base is dominated by the top tier, which means CRM performance measured in aggregate is largely measuring what happened to a small number of customers.
The top tier requires care for reasons beyond commercial value, which the protective section below addresses.
The caution about value segmentation is that it is backward-looking. A customer's historical value describes what they did, and a substantial part of CRM's job concerns customers whose future value differs from their past, particularly new ones.
Behaviour
Generally more actionable than value alone, because behaviour predicts behaviour.
The dimensions worth segmenting on.
Product preference. Sportsbook, casino, live casino, poker, bingo, and the mix. A customer who plays only slots needs different content from one who bets on football and occasionally plays live blackjack.
Session pattern. Frequency, duration and timing. A customer who plays briefly most days differs from one who plays for hours at weekends, and both differ from one who appears only around major events.
Deposit behaviour. Frequency, size, method and whether deposits are regular or clustered.
Volatility preference. Which games a casino customer chooses, since preference for high or low volatility content is stable and predicts what they will enjoy.
Bonus responsiveness. Whether a customer engages with offers, and whether their play changes when they receive them. This is directly commercially relevant, since customers who are unresponsive to offers should not be receiving expensive ones.
Channel responsiveness. Which communication channels a customer engages with, which reduces waste and reduces the irritation of contact through channels they ignore.
Seasonality. Customers whose activity concentrates around specific sports or periods, who should not be treated as lapsed in their off-season.
Lifecycle stage
The dimension most often omitted and among the most consequential, since it changes what a customer needs more than value does.
New. Registered and made a first deposit recently. They do not know the product, have no habits and are at the steepest point of the churn curve. What they need is orientation, not offers.
Establishing. Playing but not yet settled into a pattern. Habits are forming and product decisions here have disproportionate long-term effect.
Established. A stable pattern over months. They know what they like, and the questions concern depth, variety and keeping the experience from becoming stale.
Changing. Behaviour has shifted materially in either direction. Increased activity may be enthusiasm or difficulty; decreased activity may be boredom or deliberate control. Either warrants attention rather than a campaign.
Lapsing. Activity has declined towards inactivity but the customer has not clearly gone.
Lapsed. Inactive for long enough to count as churned, with the definition depending on the customer's normal pattern.
Returning. Come back after an absence, with needs different from both new and established customers.
The practical consequence is that a communication appropriate for an established customer may be entirely wrong for a new one, and operators that segment on value while ignoring lifecycle send experienced-customer content to people who joined last week.
Protective segmentation
A category that exists for reasons other than commercial value and that must take precedence over everything above.
Customers displaying risk indicators, as described in the Customer Service and Law and Compliance courses, require different treatment: exclusion from promotional contact, and in many cases active intervention rather than marketing.
Customers who have set restrictive limits, taken a time-out or been subject to an intervention should be excluded from promotional activity and from reactivation.
Customers who have self-excluded must be excluded entirely.
The critical design requirement is that these classifications override commercial segmentation rather than sitting alongside it. A customer in the top value decile who is displaying risk indicators must be treated according to the second classification, and the systems should make it impossible for a campaign targeting high-value customers to include them.
This is not merely a compliance requirement. The commercial and protective segments overlap substantially, since the customers displaying indicators are disproportionately the high-spending ones, which means the override is doing real work rather than catching edge cases.
Building it practically
Some guidance on constructing a scheme that is used.
Start from the decisions. What treatments do we actually want to differentiate? Build segments that support those rather than building a comprehensive scheme and looking for uses.
Keep it small enough to use. A dozen segments applied consistently beats sixty applied inconsistently. Every additional segment multiplies the campaign matrix and the analysis burden.
Combine dimensions deliberately. Value and lifecycle together, for example, produces a manageable grid where each cell has an obvious treatment. Combining four dimensions produces cells too small to act on.
Define membership rules explicitly and hold them stable, since segments whose definitions drift produce analysis that cannot be compared over time.
Refresh regularly. Customers move between segments, and a scheme applied at a point in time and not updated will be treating people according to what they were doing six months ago.
Check the sizes. Segments containing very few customers cannot support meaningful measurement and are usually not worth maintaining.
Review actionability annually. Which segments are actually used in targeting, and what happened to the ones that are not.
What segmentation cannot do
A closing caution.
Segmentation groups customers who are similar in the dimensions measured. It does not make them identical, and treatment designed for a segment will be wrong for some of its members.
This matters most at the extremes. The top value decile contains customers whose spending is entirely sustainable and customers for whom it is not, and no commercial segmentation distinguishes them. That distinction requires the protective classification described above, applied on behavioural indicators rather than on spend.
It also matters for new customers, where the operator has little data and segmentation is largely guesswork. The correct response to low information is usually to gather more rather than to assume, which means early-lifecycle treatment should be designed to learn about the customer rather than to target them precisely on inadequate evidence.
The general principle is that segmentation is a tool for making better decisions with limited information, not a description of who people are. Functions that treat it as the second develop confident views about customers that the data does not support.
A worked segmentation
To make this concrete, a segmentation scheme that is small enough to use and does actual work.
The primary grid combines lifecycle stage and value tier, producing a manageable set of cells.
Lifecycle across five stages: new, establishing, established, lapsing, lapsed. Value across three tiers rather than ten for operational purposes: high, mid, low, with the high tier defined narrowly enough to warrant individual attention.
That produces fifteen cells, each with a reasonably obvious treatment. New customers of any value tier receive orientation. Established high-value customers receive relationship management. Lapsing mid-value customers receive intervention. Lapsed low-value customers receive minimal investment.
Layered over this, product preference determines content rather than treatment: what a customer is shown, not how intensively they are engaged.
Behavioural flags modify treatment without creating new cells: bonus-unresponsive customers stop receiving expensive offers, channel-unresponsive customers stop receiving contact through channels they ignore, seasonal customers have their dormancy thresholds adjusted.
And protective classification overrides everything, removing customers from promotional treatment entirely regardless of which cell they occupy.
The result is a scheme with fifteen primary cells, a content dimension, a handful of modifiers and one override. It is comprehensible, it can be applied consistently, and every element changes something. Compared with a scheme of sixty segments defined on demographic and value attributes, most of which receive identical treatment, it will outperform substantially.
Segmentation and fairness
A consideration that receives little attention in this discipline and warrants some.
Segmentation determines who receives what, which means it determines who receives better terms. A customer in a high-value segment may receive enhanced offers, faster withdrawals, dedicated support and better treatment generally.
Some of this is ordinary commercial practice and is unremarkable. Some of it warrants examination.
Differential treatment on service raises a question about whether all customers receive an acceptable baseline. Faster withdrawals for high-value customers is defensible only if everyone else's withdrawals are still prompt.
Differential offers are ordinary and become a consumer protection question if the terms are materially worse for some groups without that being apparent.
Differential treatment based on inferred characteristics should be examined carefully, since a model determining who receives what may be doing so on the basis of attributes the operator would not defend if they were made explicit.
And the recurring point: enhanced treatment scaling with loss is the structure of most gambling loyalty, which is worth being clear-eyed about. Rewarding customers in proportion to what they have lost is defensible where the losses are affordable and is difficult to defend where they are not, which returns to the protective override described above.
Building segmentation without good data
A practical note, since much of this lesson assumes a data foundation many operators lack.
Where identity resolution is imperfect, attribution does not persist and costs are not attributable at customer level, sophisticated segmentation is not available. Attempting it anyway produces confident groupings built on unreliable inputs.
What is available even with weak foundations.
Recency, frequency and monetary value. The oldest segmentation approach in marketing, requiring only transaction records, and considerably more useful than nothing. When a customer last played, how often they play, and how much they have spent produces a workable grid.
Product mix, derived from gameplay records, which are usually reliable even where other data is not.
Lifecycle stage, derived from registration date and activity dates, which requires no sophistication.
Deposit behaviour, from payment records.
Communication engagement, from the email and push platforms, which is self-contained.
Those five dimensions support the great majority of useful CRM decisions, and every operator has the underlying data. An operator without advanced capability should build on them rather than waiting for a customer data platform, and should invest in the foundations in parallel.
The sequencing point matters. Operators frequently defer segmentation until the data infrastructure is ready and then discover that the infrastructure project takes two years. Starting with what is available produces value immediately and clarifies what the infrastructure actually needs to support.
Reviewing a scheme
To close, the questions to ask of any existing segmentation.
Which segments are actually used in targeting? Usually a small subset of those maintained.
Do any two segments receive identical treatment? If so, the distinction is not doing work.
When were the definitions last reviewed? Segments defined years ago may no longer describe meaningful groups.
Are customers reassigned regularly? A static assignment describes what someone was doing when it was made.
Do the segment sizes support measurement? Very small segments cannot produce reliable results.
Does the protective classification override reliably? Test it rather than assume.
Could someone new understand the scheme? Complexity that requires institutional knowledge to navigate will be applied inconsistently.
What decision would we make differently if we had a segment we do not currently have? The most useful question, because it identifies the gap rather than auditing what exists.
Segments that predict rather than describe
A refinement worth introducing, since the most valuable segmentation looks forward rather than back.
Most segmentation is descriptive: it groups customers by what they have done. Predictive segmentation groups them by what they are likely to do, which is more useful for decisions taken now.
The predictions that matter in this context.
Likelihood to churn, which identifies customers worth intervening with before they go rather than attempting winback afterwards.
Expected future value, which is more useful than historical value for deciding how much to invest in a relationship, particularly for newer customers whose history is short.
Likelihood to respond to a given offer, which prevents spending promotional value on customers who would have played anyway.
Likelihood to try a second product, which directs cross-sell effort where it might work.
Risk of harm, which is the protective classification and belongs in this list because it is a prediction like the others and is the one that matters most.
Building these requires modelling capability and data quality that many operators lack, and simplified versions are available to everyone. A churn indicator based on activity decline relative to an individual's own pattern requires no modelling and performs considerably better than a fixed threshold.
The caution is that a prediction is not a fact. A customer identified as likely to churn may not, and treating the prediction as certainty produces intervention on people who did not need it. The appropriate response to a prediction is usually a small action rather than a large one, and measuring whether the action changed anything is the only way to know whether the prediction was useful.
