Every flavourhouse product team eventually confronts the same dashboard problem: a reorder rate that looks flat across the board, and a suspicion that the flatness is an artefact of how customers are labelled. The question worth isolating is narrow and testable — when you segment repeat buyers of e-liquid by self-declared skill or experience level, does a coarse two-tier split (say, "new" versus "experienced") predict subsequent reorder behaviour better than the five-tier gradients that most loyalty platforms default to? The answer matters less for taxonomy aesthetics than for what it implies about how preference stability forms, and it points toward a counterintuitive finding: fewer tiers often carry more signal.
Why Five Tiers Feel Right and Perform Poorly
The five-tier instinct is a legacy of satisfaction research. Likert-style scales, Net Promoter bands, and the familiar bronze-to-platinum ladder all assume that finer granularity captures more variance. In survey design, that assumption has partial support — more response options can reduce ceiling effects. But reorder prediction is not a survey problem. It is a behavioural forecasting problem, and behavioural forecasting has a well-documented tolerance for coarse bins.
The mechanism is straightforward. When you collapse customers into five stated skill levels, each tier contains a mixture of people whose actual consumption patterns differ wildly from their self-description. A "Level 3" vaper might be someone who has settled into a single tobacco-adjacent profile for two years, or someone who just bought their fourth device and is about to churn through twelve fruit blends in a month. Both call themselves intermediate. Their reorder curves look nothing alike, and averaging them produces a tier mean that predicts neither.
Two tiers, by contrast, tend to split along a genuine behavioural fault line. The distinction between "still exploring" and "has a settled rotation" is not a matter of degree — it is closer to a phase change. Explorers reorder unpredictably and in small quantities; settlers reorder on a rhythm and in larger, more stable baskets. Collapsing the middle three tiers into these two poles raises within-group homogeneity even as it lowers nominal resolution.
The Variable-Ratio Trap
There is a second reason five tiers underperform, and it comes from reinforcement theory. Flavour discovery is a textbook variable-ratio schedule: most new bottles are unremarkable, a minority are genuinely rewarding, and the ratio is unknown in advance. Variable-ratio schedules are the most persistent reinforcement pattern known, which is precisely why exploration behaviour is so sticky and so poorly predicted by stated skill. A customer who labels themselves "advanced" may still be running a discovery loop, because the loop itself is rewarding independent of expertise.
Five-tier labels implicitly assume that skill level governs how much someone explores. Behaviourally, the reverse is often true: exploration governs the perception of skill. People who keep sampling new profiles describe themselves as more advanced, not because they have mastered anything, but because sampling is what they do. A two-tier split that asks about rotation stability rather than self-rated expertise sidesteps this confound.
What the Evidence Suggests About Coarse Bins
The strongest analogy comes from forecasting research rather than consumer behaviour. Philip Tetlock's long-running expert prediction tournaments, summarised in Expert Political Judgment and later in Superforecasting, found that the most accurate forecasters were not the ones with the finest-grained models. They were the ones who aggregated uncertain signals into a small number of coarse probability bins and updated those bins frequently. Granularity without reliability is noise dressed as precision.
Closer to the flavour context, the "less-is-more effect" documented by Gerd Gigerenzer and colleagues shows that simple heuristics frequently outperform more complex models when the underlying data are noisy and the sample per category is small. A five-tier segmentation of a customer base that only yields a few hundred reorder events per tier per quarter is exactly that situation. The two-tier split gives each bin enough events to produce a stable estimate; the five-tier split does not.
Kahneman's work on loss aversion adds a wrinkle worth flagging. When customers are asked to place themselves on a five-point skill scale, the middle options function as a safe harbour — few people want to claim the bottom or the top. The result is a pile-up in the middle three tiers and a distribution that is nearly useless for targeting. Forcing a binary choice ("do you have a settled rotation, yes or no") removes the safe harbour and produces a cleaner split, even though it feels cruder.
A Concrete Illustration
Consider a mid-sized flavour retailer running a reorder campaign. With five self-declared tiers, the top two tiers receive the same recommendation logic: "you're experienced, here's a complex profile." Reorder lift is modest and inconsistent. The same retailer switches to a two-tier scheme based on a single behavioural question — whether the last three orders contained at least one repeat SKU. Customers with a repeat SKU are classified as rotation-settled; those without are classified as exploring. The exploring group receives a rotating sampler; the settled group receives a larger-format repeat of their most recent SKU. Reorder rates diverge sharply between the two groups, and the divergence is stable across two subsequent quarters, whereas the five-tier version had shown no stable separation at all. The lesson is not that the sampler or the large format is magic. It is that the two-tier split sorted customers along the axis that actually drives reorder behaviour.
Design Implications for Flavour Shops
If coarse bins carry more signal, the practical question becomes which axis to bin on. Skill is a tempting axis because it is easy to ask about, but it is contaminated by the variable-ratio loop described above. Rotation stability, recency of a repeat purchase, and basket concentration are all better candidates because they are behavioural rather than declarative.
A workable two-tier scheme for most flavour shops looks like this:
- Settled: at least one repeated SKU in the last three orders, or a basket concentrated in two flavour families.
- Exploring: no repeated SKU in the last three orders, or a basket spread across four or more families.
Everything else — device type, nicotine strength, stated experience — becomes a modifier rather than a tier. This keeps the primary segmentation binary and interpretable while retaining nuance where it does not distort the prediction.
The forward-looking implication is that flavour retailers should resist the platform default. Most loyalty and CRM systems ship with five-tier ladders because those ladders sell well in software demos. They are not optimised for reorder prediction, and the evidence from forecasting and heuristic research suggests they actively degrade it. The next iteration of flavour personalisation will likely move toward fewer, behaviourally defined bins — and the shops that adopt that structure early will have a measurable edge in repeat purchase rates, not because the segmentation is sophisticated, but because it is honest about how little granularity the data can actually support.