DEEP BACKGROUND 01 OF 13 . PROPENSITY OVER DEMOGRAPHICS

How strong is the evidence that behaviour predicts better than demographics?

Strong, and the gap has been measured. In supermarket data, demographics explained seven per cent of the variation in how price sensitive a shopper was. Watching one shopping trip beat every demographic on file. Behaviour predicts behaviour.


WHY THIS IS WORTH MONEY

Most audience plans are still built on who somebody is. Age, income, postcode, household shape. The media buy follows the plan, and so does the product brief.

If who somebody is turns out to be a weak read on what they will do, the plan is not slightly off. It is pointed at the wrong object.

That is the difference between a budget that lands and one that does not, and it has been sitting in the published record for thirty years.


WHAT THE RECORD SHOWS

Five findings, and together they tell one story.


Demographics explained seven per cent. One shopping trip beat them.

Rossi, McCulloch and Allenby (Marketing Science, 1996) had scanner panel data and the question every marketer eventually asks. If you are going to send somebody a coupon, what do you actually need to know about them? So they compared knowing a household's demographics against knowing what that household had already bought.

In their words: "The demographic variables explain only 7 per cent of the variation in price sensitivity, which is a key determinant of potential profitability for target couponing."

Then the money. Measured against blanket couponing to everybody, demographics alone returned 1.12 times the gain. Observing a single purchase occasion returned about one and a half times. Short purchase histories returned two and a half times, and full history 2.55.

Read the shape of that. Everything you can buy about who a person is gets you twelve per cent better than posting to the whole list. Watching them shop once gets you fifty.


Past behaviour predicts future behaviour, and how well depends on the domain.

Ouellette and Wood (Psychological Bulletin, 1998) pooled sixty four independent studies. Across all of them, past behaviour predicted future behaviour at r = .39.

Then the finding that matters more, in their words. Past behaviour "was a weaker predictor of future behavior in domains that are encountered only annually or biannually and that present unstable contexts, r = .27... than in domains encountered on a daily or weekly basis and that present stable contexts, r = .59".

So behaviour predicts behaviour, but not everywhere equally. In a domain somebody lives in weekly, the past is a strong read on the future. In a domain they touch twice a year, it is a weak one.

That is a constraint and it belongs on the page rather than in a footnote. It is also, stated the other way round, the condition the Cone was already built on, because a Cone is placed on a domain people live in and not on an occasional purchase.


What people say they will do is a weak stand-in for what they do.

Sheeran and Webb (Social and Personality Psychology Compass, 2016) put the number on it. Intentions and later behaviour correlate at r+ = .53, which sounds reassuring until you look at what happens when somebody's intention is actually changed. Across the experimental evidence, "a medium-to-large-sized change in intentions led to only a small-to-medium-sized change in behavior (d+ = .36)".

Change what a person intends and most of the behaviour stays exactly where it was. This is the paper to reach for when a room wants to run a survey asking people whether they would buy the thing.


And the demographic data being bought is often not accurate.

Neumann, Tucker and Whitfield (Marketing Science, 2019) bought audience segments from third party data vendors and checked them against known truth.

Across fourteen vendors, gender accuracy averaged 42.3 per cent, ranging from 25.7 to 62.7. A coin toss is fifty. Age was worse. Identifying 18 to 24 year olds averaged 10.7 per cent. Identifying men aged 25 to 54 averaged 24.4 per cent, against a natural population rate of 26.5 per cent, so paying for that segment did slightly worse than picking at random.

Interest based segments held up far better in the same tests. Sports averaged 87.4 per cent, fitness 82.1.

That pair is the whole argument sitting inside one dataset. The attributes describing who a person is were unreliable. The attributes describing what a person does were reliable.


The man who put segmentation into business language said it himself.

Daniel Yankelovich wrote New Criteria for Market Segmentation in Harvard Business Review in March 1964, the article that put non-demographic segmentation into general business use. Forty two years later he came back to it, with David Meer, in the same journal.

"Psychographics may capture some truth about real people's lifestyles, attitudes, self-image, and aspirations, but it is very weak at predicting what any of these people is likely to purchase in any given product category."

And: "While relevant attitudes, values, and expressed preferences can bring color and insight to a segmentation, they lack the predictive power of actual purchase behavior, such as heaviness of use, brand switching, and retail-format or channel selection."

The person with the strongest claim to have invented the practice spent his return to it saying that what people do predicts, and what people say about themselves does not.


THE STRONGEST OBJECTION, AND THE ANSWER TO IT

Byron Sharp and the Ehrenberg-Bass Institute. If brands inside a category share their buyers, and the buyers of competing brands look the same as each other, then carving an audience up is a waste of money and the answer is broad reach.

The evidence under that is real and it is recent. Anesbury, Winchester and Kennedy (Marketing Letters, 2017) looked at seven hundred brands, more than sixty packaged goods categories and more than one hundred and sixty variables. Competing brands' user profiles seldom differ from one another, and seldom change across three to six years.

Two things are true about that.

The first is that it is a finding this page rests on rather than one that damages it. What they show is that demographics do not separate the buyers of one brand from the buyers of its rival. That is our claim, arrived at from the other side of the table.

The second is that their conclusion is about brands inside a category. A Cone is not placed on a brand. It is placed on the domain the brand sits inside, and it reads how much of a person's life that domain occupies. Different object, different question.

Where it is not settled: Ehrenberg-Bass would say that once you know profiles do not differ, the growth answer is reach and availability rather than any deeper read of the audience. That is a live disagreement and this page does not resolve it. The claim here is narrower and holds either way. If you are going to describe an audience at all, describe it by what it does.


WHERE THE EVIDENCE STOPS

Rossi and colleagues is 1996, one product category, United States scanner panel data. It is strong on price sensitivity and couponing. It is not a general law of everything.

Ouellette and Wood is a constraint as much as a support. In domains people encounter once or twice a year, a behavioural read is weak at r = .27. Say that out loud before anybody commissions a Cone on an annual purchase.

Sheeran and Webb tested changing intention and measuring behaviour. It does not say intention is worthless. It says intention is a weak lever.

Neumann and colleagues tested United States third party vendors in one market at one time. Vendor quality is not a constant and the figures should not be quoted as one.

None of these papers tests the Cone. They test the ground the Cone stands on. That distinction stays on the page.


WHERE THE CONE SITS

Propensity is what a person is already doing, which is how you can see where more of it sits. Not what they say, not what they intend, and not who they are on a form.

The four behaviours, Fanatics, Enthusiasts, Casuals and Indifferents, are positions inside one domain. The same person sits at a different position in every other domain of their life, which is why a behavioural read has to name the domain before it names anybody.

The literature above says three things, and the Cone was built on all three. Behaviour beats demographics as a read. Stated intention is a weak stand-in for behaviour. A behavioural read is strongest where the domain is lived in regularly.

And the rule that governs what happens next: it is the idea that moves through the cone, not the people, moved up it. The job is to meet each position where it already is, so its propensity is matched, respected and employed.


WHAT TO DO ON MONDAY

Take the audience data you already hold. Sort it once by demographic and once by depth and frequency of engagement in your domain.

Put the two sorts beside each other and ask which one produces groups that behave differently from one another. You do not need to buy anything to run that.

If the demographic sort produces four groups that behave the same, you have just found where the money is going.


Propensity is pillar one of thirteen. The instrument built on it is the Cone, and how it works is the next page.


SOURCES

Rossi, P. E., McCulloch, R. E., & Allenby, G. M. (1996). The value of purchase history data in target marketing. Marketing Science, 15(4), 321-340. doi:10.1287/mksc.15.4.321. Full text: rob-mcculloch.org/some_papers_and_talks/papers/published/1996_The_Value_of.pdf
Ouellette, J. A., & Wood, W. (1998). Habit and intention in everyday life: the multiple processes by which past behavior predicts future behavior. Psychological Bulletin, 124(1), 54-74. Full text: dornsife.usc.edu/wendy-wood (Ouellette.Wood_.1998 PDF)
Sheeran, P., & Webb, T. L. (2016). The intention-behavior gap. Social and Personality Psychology Compass, 10(9), 503-518. doi:10.1111/spc3.12265. Full text: eprints.whiterose.ac.uk/107519
Neumann, N., Tucker, C. E., & Whitfield, T. (2019). Frontiers: how effective is third-party consumer profiling? Evidence from field studies. Marketing Science, 38(6), 918-926. doi:10.1287/mksc.2019.1188. Vendor tables: ide.mit.edu/wp-content/uploads/2022/02/c.tucker.pdf
Yankelovich, D., & Meer, D. (2006). Rediscovering market segmentation. Harvard Business Review, February 2006. hbr.org/2006/02/rediscovering-market-segmentation
Yankelovich, D. (1964). New criteria for market segmentation. Harvard Business Review, March 1964. hbr.org/1964/03/new-criteria-for-market-segmentation
Anesbury, Z., Winchester, M., & Kennedy, R. (2017). Brand user profiles seldom change and seldom differ. Marketing Letters, 28(4), 523-535. doi:10.1007/s11002-017-9437-2

Superfandom . Deep Background . https://superfandom.ai/science/propensity-over-demographics/
Matt Hart . published 2026-09-08 . last changed 2026-09-09