Social proof

Evidence

Social proof is an umbrella term. Research draws a sharper line. A descriptive norm states what others actually do (“most customers pay by invoice”). An injunctive norm states what one ought to do (“please sort your waste”). The two behave differently, and part of the contradictory evidence in this field comes from mixing them up.

The best-known field study is by Goldstein, Cialdini and Griskevicius (2008). It is the most-cited single study in this field. In the rooms of a single hotel in the US, a sign asked guests to reuse their towels. In the first experiment the descriptive-norm version produced 44.1 percent reuse against 35.1 percent for the usual environmental appeal (N = 433, Φ = 0.09). The Φ measures the strength of the association between two yes-no features and runs from 0 to 1; values around 0.1 count as small. In the second experiment the strongest version was the one referring to guests in that same room: 49.3 percent against 42.8 percent for the three other normative versions combined (N = 1,318, Φ = 0.06).

Two things are rarely quoted alongside it. First, these are small effects: Φ = 0.06 to 0.09 amounts to differences of six to nine percentage points. Second, the room version worked even though respondents rated “guest in this room” the least important of the group memberships tested — less important than gender or citizenship, and those two did no better than the general guest norm. The rule of thumb is therefore not “similar people” but the same situation.

Schultz and colleagues (2007) had shown the other side a year earlier. In one Californian town, 287 households had their electricity use compared with the neighbourhood average and hung on their door. Households above the average then saved 1.22 kilowatt-hours a day (d = 0.55). Households below it used 0.89 kilowatt-hours more (d = 0.52) — the norm pulled them up. When the feedback also carried an evaluative mark, that increase was no longer detectable. The mark was a hand-drawn smiley: a happy one for frugal households, a sad one for the heavy consumers.

After four weeks the balance reversed. The boomerang among the frugal households held (d = 0.59), while the saving among the heavy consumers was no longer significant without the evaluative mark. The destructive half of the effect therefore outlasted the constructive half.

Melnyk, Carrillat and Melnyk (2022) pool 250 effect sizes from 136 papers covering 112,478 respondents in 22 countries. The mean sits at r = 0.254 — the r measures how closely two quantities go together and runs from 0 to 1. Descriptive norms work markedly better there than injunctive ones (r = 0.305 against 0.223). That supports the direction — with one qualification: only 40 of the 250 effect sizes come from experiments. The large remainder are associations between a perceived norm and reported behaviour, and only about ten percent of the papers come from marketing.

The second meta-analysis is larger, and it says something awkward. Bergquist, Nilsson and Schultz (2019) analyse 91 field experiments from 74 papers covering 227,730 observations, published between 1982 and 2019 across 18 countries. Against control conditions, norms there work at d = 0.32 (95 percent interval 0.28 to 0.37). Measured against what other environmental interventions achieve, the authors call that a medium-sized effect.

What matters is the route by which the norm arrives. The analysis separates norms spoken as a message from norms that can only be read off the surroundings — a clean pavement, lights left off, a nearly empty cookie jar.

The difference between the two rows is clear (p < .001). A sign, a letter, a sentence on a page — anything that can be worded — falls into the lower row, and 0.18 sits below what counts as a small effect. The authors put the gap down to a spoken norm having a sender, which makes it recognisable as an attempt to influence; a setting has none.

Two side findings from the same paper. The expectation that norms would work more strongly in collectivist cultures came out reversed: the association with Hofstede’s individualism index is positive (r = .23), so norms worked more strongly in more individualistic countries. And studies with students arrive at d = 0.59, studies with other samples at d = 0.25 — with the student studies almost all run on a campus.

Goldstein’s towel sign sits near the bottom of this collection. Experiment 1 is listed at d = 0.19, whose interval begins at exactly 0.00, experiment 2 at d = 0.11. The two German repetitions by Bohner and Schlüter are recorded at d = −0.06 and d = −0.80, that is, with the sign reversed.

Eleven years on, the same authors published a retrospective on their own study (Schultz et al., 2018) reporting what became of the finding in the field. Since 2008 the company Opower has been sending out electricity bills carrying exactly this feedback — a neighbourhood comparison plus an evaluative mark — by now to more than 60 million households. Every rollout is built as a randomised trial: at a utility serving 100,000 homes, half receive the reports and half do not. Across more than a hundred such rollouts the savings run between 2 and 5 percent. After the reports stopped the effect halved, but stayed measurable against a control group that never received any (Allcott & Rogers, 2014).

The boomerang showed up at scale too. A reanalysis of the Opower data (Costa & Kahn, 2013) found a group whose consumption rose in response to the message: politically conservative households above the norm who did not donate to environmental organisations and did not pay for renewable energy. The evaluative mark does not catch the boomerang for everyone. The authors put this down to the referent group: someone who does not identify with the neighbours is not pulled by their norm but pushed away.

The distinction itself comes from a series of field studies in 1990. Cialdini, Reno and Kallgren coined the terms descriptive and injunctive across five experiments on littering. Their claim: norms work only while attention rests on them — and only the norm attention rests on. In a littered parking garage 41 percent dropped a handbill, in a clean one 11. When a confederate drew attention to the floor, both effects grew: 54 against 6 percent.

The best-known finding from it runs against intuition. A single piece of litter in an otherwise clean setting pushed littering below the level of the clean setting — it makes visible that nobody drops anything here.

And the boomerang is already here, seventeen years before Schultz. The authors record themselves that a descriptive norm in a heavily littered setting increases littering — “hardly a desirable outcome for any but theory-testing purposes”. Their remedy is the same as Schultz’s: in experiment 4 the same litter lay either scattered or swept into piles. A swept pile is a mark of disapproval, that is, an injunctive norm. Under attention-focusing, 45 percent littered where it was scattered but only 18 percent where it was swept.

The same-room rule has been tested directly in Germany — and it failed. Bohner and Schlüter (2014) repeated Goldstein’s setup in two German hotels, with 724 and 204 observed guests and a Latin square across the floors. Three results run against the original.

First, reuse in Germany sits at an entirely different level: even the condition with no message at all reached 64.3 percent, above every condition in Goldstein. Second, no normative version beat the standard environmental appeal (81.9 against 83.7 percent, p = .62); in the second study the environmental appeal was actually superior to the normative ones (93.3 against 76.5 percent, p = .039). Third, the same-room advantage significantly reversed: the room version worked worse than the general guest norm (78.0 against 85.6 percent, p = .017).

Two further tests point the same way. Schultz and colleagues (2008) found in the United States that a purely descriptive norm did not raise reuse against a control group at all — only the combination with an injunctive norm worked; and in their data the general norm came out ahead of the room norm. Reese and colleagues (2014) in Austria and Switzerland did find fewer towels in the same-room condition, but no advantage of any normative version over the environmental appeal. Both papers are reported here at second hand, through Bohner and Schlüter.

An independent team finds the norm message ineffective before it can even work. Greene and colleagues (2025) tested nine kinds of message on 900 and 599 respondents and then in the field at two Australian hotels. The norm message failed its own manipulation check: it did not significantly raise normative belief. And normative belief was the only one of the eight constructs tested that correlated with neither intention measure (r = .146 and .012, both non-significant). What did work in the field was a message about effort. The authors also record that Goldstein and colleagues never tested whether their message strengthened normative belief at all.

Limits

The boomerang is not an edge case. Every descriptive norm also names a target value that people approach from above and from below. A shop writing “most customers choose the middle plan” may thereby pull customers away from the expensive one.

A minority norm works against its own intent. “Only a few people use our app so far” reliably says that it is not the done thing. Figures you might want to state because they are honest can undercut the message.

The solid field data come mostly from energy and environmental contexts, not from shops. Towels and electricity bills are recurring routines with a visible benchmark. Whether the same mechanism carries a one-off purchase decision remains open. The nearest available tests in Melnyk et al. (2022) argue against the worry rather than for it: whether a behaviour is public or private, costs much or little effort, is demanded once or permanently — none of those differences changed the effect there. None of them captures a one-off purchase exactly, but together they give the worry little support.

Invented numbers are not a design choice. Claims such as “1,284 people are viewing this offer right now” without a real basis are misleading; whether they are unlawful in a given case belongs in a professional legal review. In design terms the rule is simple: a figure whose derivation you could not explain to a customer does not belong on the page.

Cialdini and colleagues could not check their own manipulation. Because all five studies ran in the field, no manipulation checks were possible. The authors concede that one may be less confident that the intended type of norm actually did the work — the single piece of litter might just as well have activated the injunctive norm.

Many of the individual comparisons do not reach significance. In experiment 1 neither effect in the clean setting was significant on its own, in experiment 2 not the difference between zero and one piece of litter, in experiment 4 neither direction separately. What is significant are the predicted patterns across conditions.

The figures on reach are second hand. What is reported about Opower comes from the authors’ own retrospective on their own line of work, written for a special issue about their most-cited paper. One of the two sources for the 2-to-5-percent range is a TED talk. It serves as an order of magnitude, not as a dependable value.

Which figure makes a strong norm is open. The 2007 study gave a mean; other work gives percentages — “75% of guests who stay in this hotel …”. In 2018 the authors state that nobody knows which presentation yields a strong descriptive norm. Early work suggests a low but rapidly rising figure can motivate while a high but falling one does not; that field, they write, is just beginning.

The 2007 sample was narrowly drawn. The review board required an opt-out protocol, so only households whose electricity meter was visible from the street and easily accessible were included. The team read the meters themselves, because the utility withheld the consumption data citing proprietary data and customer confidentiality.

The large meta-analysis is highly heterogeneous. In Bergquist et al. (2019), 88 percent of the differences between studies are not down to chance but to real differences between them. The check for publication bias comes out asymmetric: sixteen studies would have to be added for the picture to balance. The authors name three possible reasons and settle on none of them.

That paper’s most important finding is not cleanly isolated. The authors record themselves that the studies using norms read off the surroundings mostly examined littering — so the mode of conveyance and the behavioural domain coincide. A meta-analysis is an observational study, and only a single one of the included papers tested the two modes directly against each other.

The key study has been directly replicated and failed on two of three points. Bohner and Schlüter (2014) found: the room norm reversed, and no normative version beat the standard environmental appeal. What was confirmed is only that any message works better than none.

What carries the effect is the meta-analysis by Melnyk and colleagues across 250 effect sizes — not the towel sign. The best-known single study in the field is thus also the worst-confirmed.