Behavioural Profiling

How accurate is claimed purchase frequency in surveys?

Less accurate than it looks, and biased rather than merely imprecise. People answer by applying a rate heuristic ("about twice a month") or by recalling instances and extrapolating, so regular behaviours are over-reported, irregular ones are under-reported, socially approved behaviours are inflated, and events from before the window are pulled into it. All four effects worsen as the recall window lengthens, so a twelve-month frequency claim carries much less information than a seven-day one. The most important operational rule is never to multiply a claimed frequency into an annual or market volume, because that compounds the bias into a number that then enters a business case as a fact.

What do I do when survey answers and transactional data disagree?

Measure the gap rather than choosing a side. Cross-tabulate claimed against observed where the data links at individual level, and report the size, direction and pattern of the difference. The gap is usually patterned: heavier users over-estimate more, light users less, so the two errors together compress the true distribution towards the middle. Treat that as a finding, because it tells you how the audience understands its own behaviour, which is directly relevant to how they will respond to any communication about it. Do not average the two sources, and do not silently prefer the survey because it has more variables attached to it.

Should I measure brand loyalty?

Usually not in the form it is normally asked. In most repeat-purchase categories buyers hold a repertoire of several brands and allocate volume across it, and exclusive buyers are a minority even among a brand's own customers. A self-reported "main brand" question therefore measures self-image rather than behaviour. Report repertoire size, share of category requirements and penetration instead. Penetration and share of requirements separate two different growth routes, more buyers or more share from existing buyers, which a single loyalty number hides.

Why is profiling the heavy user a mistake?

Because heavy users are unrepresentative by construction, and in most measured categories the large majority of buyers buy infrequently and collectively account for a substantial share of volume, often more than the heavy minority. A proposition designed around heavy users addresses the people least in need of persuading, and the arithmetic of growth usually favours reaching more light buyers. Report the full buyer distribution, showing share of buyers and share of volume by frequency band, before any heavy-user content appears. Heavy-user analysis is legitimate; it is a description of a minority and should be framed as one.

Should I profile people or occasions?

Test it rather than assuming. Where occasion-level data exists, compare within-person variation with between-person variation on the key behaviour. In many categories the same individual behaves quite differently at different times, and the within-person variation is larger, which means a person-level profile produces an average that fits no actual situation and a segmentation that will not stabilise. Where that is the case, profile occasions and describe people by their occasion repertoire. Where occasion data does not exist, say that the unit could not be tested.

Does an attitude-behaviour correlation show that attitude drives behaviour?

No, and in behavioural research the reverse direction is not a technicality. People report warmer attitudes towards brands they already use, so a correlation between brand attitude and purchase frequency is at least as consistent with usage producing attitude as the other way round. Report the association with its base, state that the direction is not established by a cross-sectional design, and name the design that would settle it if the answer matters to the decision.

Behaviour changed between waves. Did people change?

You cannot tell from cross-sectional data. Three explanations compete: the same people changed, different people are in the sample, or the same people are in different circumstances such as a price change, a supply problem or a seasonal effect. Only panel data on the same individuals separates the first from the second. Where you have it, decompose the change into buyers gained and lost versus rate changes among retained buyers; that decomposition is usually more actionable than the headline movement.

What are the blind spots of transactional data?

Coverage, not accuracy. A single retailer's file describes purchases at that retailer; an app's telemetry describes use of that app; a card record misses cash. It also usually resolves to a household or an account rather than a person, which inflates every per-person frequency, and it carries no context about why anything happened. State the coverage boundary on every figure derived from it, and never infer that a customer buys nowhere else from the absence of competitor records.

How should behavioural groups be derived?

From recorded behaviour: frequency, recency, repertoire, share of category, occasion mix, channel mix. Not from a self-classification item, which measures self-image. Then check persistence: someone in the heaviest band this quarter is frequently in a middle band next quarter, purely through regression to the mean. A behavioural group that does not persist across periods describes a period rather than a group of people, and it will not support a targeting strategy.

Can I build a behavioural profile from survey data alone?

Yes, provided it is honest about what it is. Label every figure as claimed, state the likely direction of bias for each measure, use the shortest recall window available, report distributions rather than means on skewed behaviours, and calculate no derived volumes from it. Confidence caps at moderate for relative statements and low for absolute levels. Then name the observed source that would fix it, because that sentence is often what gets the data acquired for the next study.