LESSON
0.01

What people say they'll do isn't what they do

The oldest open wound in behavioural science: the gap between stated intention and real action, and why it should change how you read every survey.

WRITTEN BY
Mike Popesku
PUBLISHED
September 6, 2026

What the science says

Consensus

"Behaviour" sounds simple, but the discipline splits it in two. Revealed behaviour is what you actually choose under real constraints: the purchase, the vote cast, the click. Economists formalised this as revealed preference (Samuelson, 1938). Reported behaviour is what you tell a researcher: your intention, your attitude, your recalled action. It's what most surveys actually collect. The gap between the two is well-established and large.

The canonical demonstration is LaPiere (1934). Travelling the US with a Chinese couple, he was turned away by just one of about 250 establishments. Yet when he later wrote to those same venues, over 90% said they would refuse Chinese guests. Wicker's (1969) review found attitude-behaviour correlations were typically weak. Modern work pinned the number down. Sheeran's (2002) synthesis of 422 studies put the intention-behaviour correlation at r ≈ 0.53, about 28% of the variance, and Webb & Sheeran's (2006) meta-analysis of 47 experiments delivered the decisive causal point: a medium-to-large change in intention (d = 0.66) produced only a small-to-medium change in behaviour (d = 0.36). The citation record backs this strongly. These findings are overwhelmingly supported and rarely contradicted.

Controversies

The gap itself is not in dispute. What's contested is how to model and close it. The dominant framework, the Theory of Planned Behaviour (Ajzen, 1991), inserts intention as the bridge between attitudes/norms and action, and it is among the most-cited papers in all of social science. But a substantial critical strand argues it explains behaviour after the fact better than it predicts it, assumes deliberate reasoning where much behaviour is habitual or cued, and still leaves most of the variance unexplained. So the picture is a near-universal framework that a growing literature treats with suspicion.

Limitations

Two cautions sit underneath all of this. First, a self-report is itself a behaviour: an answer produced for an interviewer, shaped by social desirability, faulty recall, and the demands of the question. Think about it: when you answer a question in a survey you are also subconsciously thinking about how you will be received, if anyone will have an opinion on you, "oooh what would they say if they knew this is how I reply" or simply wanting to project themselves as holier or more ethical, in other words "better" than they would actually behave. That's the thinking that makes self reporting an interesting behaviour to be mindful of to begin with.

Second, revealed-preference data is silent on the "why": you see the choice, not the reasoning or the option not taken. And the size of the say-do gap varies. It's smaller for easy, deliberate, well-formed intentions and larger for effortful or habitual behaviour. This is so typical in digital fields. Think SaaS companies for example. They are inundated with behavioural data from their users. They know how much time their users spent on each display of their app, where they clicked, when, how many times, absolutely everything historical. They have zero proof on the underlying reasons they behaved that way though. And even worse, they have zero understanding on the things the user didn't do, which is another set itself.

Open questions

We still don't have a clean account of when stated intention predicts action well versus badly, and whether the known gap-closers (such as forming specific "if-then" implementation plans) hold up reliably at scale and over time. This is where behavioural sciences, market research, marketing strategy, etc... stop becoming predictable practices and become fuzzier. It's all because of human consciousness. At some point, you have to learn to live with a certain degree of uncertainty and the need to do your own interpretation of all the data you might have collected, both self-reported and revealed.

"So what?"

Most commercial and political research runs on reported behaviour: surveys, intent scales, declared preference. The science says that data is a weak proxy for what people will actually do. If you treat a stated intention as a promise, you will routinely over-forecast. The discipline is to demote self-report to a hypothesis and check it against behaviour.

For companies

Purchase-intent surveys systematically overstate. "Definitely would buy" is not actual demand. It's a declared attitude with a loose relationship to the till. Before you greenlight on intent, calibrate those scores against base rates and real behavioural data: scanner panels, A/B tests, conjoint share-of-preference, actual repeat rates. The number to trust is the one tied to a behaviour.

For political parties

"Will you vote?" overstates turnout, and declared preference is soft. The people who intend to vote but don't are the say-do gap in person. Weight stated polling against behavioural signals (past turnout, registration, early-vote data), and treat persuadable-but-inactive supporters as a follow-through problem. Reduce the friction of acting, rather than only trying to persuade.

For government

Stated support for a policy is not uptake or compliance. Survey approval of a scheme tells you little about whether people will enrol, switch, or change habits. Design for the action itself (defaults, simple steps, removing barriers), and measure the behaviour, not the declared approval.

How to use this

One rule: triangulate, and never assume the sources converge. Put stated data (surveys), observed data (experiments, watching), and trace data (transactions, logs, records) side by side. When they agree, you have a finding. When they disagree, which is often, the behaviour wins, and the self-report becomes a question to investigate rather than an answer to act on.

Case studies

References

  1. LaPiere (1934), Social Forces. 10.2307/2570339
  2. Wicker (1969), J. Social Issues. 10.1111/j.1540-4560.1969.tb00619.x
  3. Samuelson (1938), Economica. 10.2307/2548836
  4. Ajzen (1991), OBHDP. 10.1016/0749-5978(91)90020-T
  5. Sheeran (2002), Eur. Rev. Soc. Psychol.. 10.1080/14792772143000003
  6. Webb & Sheeran (2006), Psychological Bulletin. 10.1037/0033-2909.132.2.249
  7. Sheeran & Webb (2016), Soc. Personal. Psychol. Compass. 10.1111/spc3.12265