Anchoring: Why the First Number Wins

That split is unusual and worth taking seriously. The core effect is among the most robust results in judgement research, and the commercially interesting extensions of it are among the least.

The original demonstration

Amos Tversky and Daniel Kahneman described anchoring in 1974. Participants watched a wheel of fortune stop on a number, then estimated the percentage of African countries in the United Nations.

Those who saw a high number gave substantially higher estimates than those who saw a low one. The wheel was visibly random and had nothing to do with the question.

The proposed mechanism was adjustment from a starting point — you begin at the anchor and move away from it, and you stop moving too soon. Later work has proposed selective accessibility instead: the anchor makes anchor-consistent information more available in memory, which biases the estimate without any deliberate adjustment happening at all. The mechanism is still argued over.

It survived the replication crisis

This is worth stating plainly, because this site spends a lot of time on findings that did not.

The Many Labs project, which tested thirteen classic effects across 36 samples, found anchoring effects that were stronger than the original studies reported. A large meta-analysis pooling the available data found a substantial average effect across hundreds of studies and tens of thousands of participants.

So when anchoring is described as one of the most robust phenomena in judgement and decision-making, that is accurate — for the classic form, where a person consciously considers a number before making a numeric estimate.

Before the exceptions, one clarification about scale. The meta-analytic estimate covers hundreds of studies and many thousands of participants, and the average effect is large by the standards of this field. Heterogeneity between studies is also high, which means the size of the effect varies considerably with the task rather than being a single fixed quantity.

What does not replicate

Three extensions are widely repeated and considerably weaker than the core effect.

Subliminal and incidental anchors

The claim that numbers you never consciously registered — a room number, a jersey number, a price glimpsed in passing — still move your estimates has not held up. Two high-powered preregistered replications of subliminal anchoring found no evidence of it, and a large replication of incidental anchoring produced an effect close to zero.

The pattern suggests something specific: deliberately considering the anchor as a possible value appears to be necessary. An anchor you did not notice is not doing work.

Anchors known to be random

This one is awkward, because it undercuts the most famous demonstration. The meta-analytic work found that anchors explicitly communicated as randomly determined did not produce effects.

The wheel-of-fortune study is exactly that design. The classic effect is real; the classic demonstration of it may be the weakest case in the family.

Willingness to pay

Anchoring on what people will pay for something — the version that matters commercially — has produced inconsistent results. A well-known study in which participants wrote down digits from their social security number before bidding on goods failed direct replication, and several other valuation studies have found no effect.

Numeric estimates of facts anchor reliably. What someone will actually hand over money for does not, at least not dependably.

Why the mechanism matters for the exceptions

The two leading explanations make different predictions, and the replication failures help distinguish them.

Insufficient adjustment says you start at the anchor and move away too little. That requires you to notice the anchor and treat it as a starting point — which predicts exactly what was found: unnoticed anchors do nothing.

Selective accessibility says the anchor makes consistent information easier to retrieve, biasing the estimate without deliberate adjustment. That would predict effects from incidental numbers, and those are the ones that failed.

The pattern in the replication record therefore favours an account requiring conscious consideration, which is a rare case of failed replications clarifying a mechanism rather than merely deleting a finding.

Where it does operate

The applications with support are mostly ones where a real number is deliberately presented and considered.

A list price beside a sale price. An opening offer in a negotiation. A suggested donation amount. A default quantity. A first salary figure. In each case the number is explicit, relevant-seeming, and consciously processed — the conditions under which anchoring is strongest.

The practical defence follows from the mechanism: generate your own number before you see theirs. An estimate formed independently has nothing to adjust away from, and anchoring has no starting point to supply.

It is instructive to compare this with the decoy effect, which is used in almost identical commercial advice and which fared very differently under the same scrutiny.

One practical asymmetry is worth naming. In a negotiation, the first number is disproportionately powerful — but only if the other side engages with it rather than dismissing it outright. An opening figure treated as absurd and set aside functions less like an anchor and more like noise, which is consistent with the finding that anchors known to be arbitrary do little.

What to take from it

Anchoring is a genuine, large, repeatedly confirmed effect on numeric estimates — one of the better-established findings in the field.

It is also routinely oversold. The version in sales training, where any number in the environment steers a buyer without their noticing, is not supported. The version in the research, where a consciously considered number shifts a subsequent estimate, is.

That distinction is the whole practical difference, and it is the same shape as several other entries here: a real finding, a stronger claim built on top of it, and only one of the two surviving examination — see also what actually remains of the Dunning-Kruger effect.