Anchoring effect

The hopeful asking price

A used road bike is listed at $900. The buyer has already checked a dozen recent sales of the same model, which went for $400 to $500. They offer $700, settle at $750, and leave pleased with the discount.

The buyer had better evidence about the bike’s value than the seller’s number: the actual sale prices. Measured against those, $750 is well above the market. The listing price set the scale the buyer bargained on, and “a discount from $900” replaced “a fair price for this bike”.

The number in the question

In a planning meeting, a new intern who has never seen the quarterly report asks: “Could we get it done in under two days?” Then everyone writes down their own estimate. The answers cluster around two to three days, although the last three reports each took about a week.

This is the less obvious form, and it’s the one studied most in the lab: a comparison question (“more or less than X?”) followed by an estimate. Nobody claimed two days was right; it was a guess inside a question. But answering the question means thinking about what would make two days possible, and the estimates start from there instead of from the track record.

Starting from the right place, stopping too soon

A volunteer coordinator plans supplies for this year’s neighborhood cleanup. Last year 60 people registered in advance and 200 came. This year 180 have registered, and she orders supplies for 240.

Last year’s turnout is a sensible starting point. The error is the size of the adjustment: registrations have tripled, which points well above 200, but the estimate barely moves from where it began. An anchor doesn’t have to be irrelevant to mislead; holding on to a relevant one too tightly is the same error.

• Comparison anchoring: the standard version. A “higher or lower than X?” question, then an estimate. This is the form that replicates most strongly (see Evidence).
• Partial-computation anchoring: the anchor comes from your own first steps. In a 1974 demonstration, high school students given five seconds to estimate 8×7×6×5×4×3×2×1 gave a median of 2,250; those shown 1×2×3×4×5×6×7×8 gave 512. The correct answer is 40,320. Both groups extrapolated from the first few multiplications, and both fell far short.
• Insufficient adjustment from a relevant value: last year’s figure, a previous estimate or a default setting, moved less than new evidence warrants (the third example above).

• When the starting number is evidence. An appraiser’s estimate, a price guide or recent comparable sales all carry information about the answer. Moving toward them is updating, as long as they get the weight their source deserves.
• When you adjust as far as the evidence says. Starting from a known reference point (last year, a similar case) is often the best available method. It becomes anchoring only when the adjustment stops short of what the new information supports.
• When a rough figure is all you have. Using it is reasonable if you hold it loosely and revise it once real information arrives.

The test: does this number tell me something about the answer, and would my estimate change if I had seen a very different number first?

The auction catalog estimate

A first-time bidder at a furniture auction has no idea what a mid-century sideboard is worth. The catalog lists an estimate of $1,200–$1,500, and she decides her limit is $1,400.

Her figure is close to the number she saw, but the number isn’t arbitrary: it was set by a specialist who studied the piece and recent sales. She has no better information, so relying on it is using the starting number as evidence. Compare the bike buyer above, who already had better evidence and let a weaker number override it.

Last year’s budget, rebuilt line by line

A club treasurer starts next year’s budget from this year’s $8,000. The hall rent rises $600, one event is dropped (saving $900) and insurance renewal quotes come in $150 higher, so she proposes $7,850.

The budget starts from an existing figure and ends close to it, but each change is driven by a specific piece of evidence, and each is adjusted in full. That’s adjusting as far as the evidence says. It would be anchoring if she had nudged the total by a round amount without costing the changes.

Tversky and Kahneman’s original account was insufficient adjustment: people begin at the anchor and adjust until they reach a plausible value, then stop, which leaves the estimate on the anchor’s side of the range.

A second account, developed by Thomas Mussweiler and Fritz Strack, is selective accessibility. Considering whether the answer could be the anchor brings to mind facts consistent with that possibility, and those facts then shape the estimate. It works like testing an idea by looking for reasons it’s true: the anchor becomes a hypothesis, and the search that follows favors it. The two accounts aren’t mutually exclusive, and researchers still debate how much each explains.

• Argue against the number. In a study by Mussweiler, Strack and Pfeiffer (2000), 60 car mechanics and dealers priced a ten-year-old car after hearing a high or a low figure. Without counterarguments, the high-anchor group’s prices were clearly higher than the low-anchor group’s. Among those asked to list reasons the figure was wrong, the gap was smaller and only marginally significant. It was a single small study, but a 2026 meta-analysis likewise associates debiasing interventions with smaller anchoring effects.
• Ask what the number is based on. If it comes from someone who knows the answer, weigh it as evidence. If it doesn’t, the task is to set it aside and build the estimate from what you know.
• Don’t count on incentives alone. Tversky and Kahneman reported that paying for accuracy did not reduce anchoring in their demonstration; the recent meta-analysis finds incentives associated with smaller effects, so the picture on incentives is mixed.

Status: replicates robustly for the standard paradigm, in which people consider an anchor before estimating. A multi-lab, preregistered replication and a large meta-analysis that corrects for publication bias both support it. Anchors that are merely present in the environment are a different matter (see below).

• Tversky and Kahneman (1974) described anchoring as one of three judgment heuristics. People saw a number between 0 and 100 from a spun wheel of fortune, said whether a quantity was higher or lower, then estimated it. Median estimates of the percentage of African countries in the United Nations were 25 for those who received 10, and 45 for those who received 65. This wheel-of-fortune version is the famous one, but it isn’t the one the large replications below reran, and the recent meta-analysis finds weaker effects for anchors drawn from random numbers.
• Many Labs 1 (Klein et al., 2014) reran 13 effects under one preregistered protocol in 36 samples, 25 in the US and 11 elsewhere (about 6,300 participants). Its anchoring items came from Jacowitz and Kahneman (1995): participants saw a high or a low figure, then estimated the distance from San Francisco to New York City, the population of Chicago, the height of Mount Everest and the number of babies born per day in the US. All four replicated in all 36 samples, with large pooled effects (d of roughly 1.2 to 2.4, depending on the question). The size varied between samples, which the authors suggest may reflect differences in what people already knew about the quantities.
• Schley and Weingarten (2026) meta-analyzed 2,601 effect sizes and found a large overall effect (g = 0.83) that stays large after accounting for extensive publication bias. They report reduced or null effects for incidental anchors, anchors on a different dimension from the estimate, anchors taken from random numbers, and when incentives or debiasing interventions are present.
• Incidental anchors did not replicate. Critcher and Gilovich (2008) reported that a phone named “P97” led to higher sales estimates than one named “P17”. Many Labs 2 (Klein et al., 2018) reran it with 6,826 participants and found an effect indistinguishable from zero (d = 0.04, against 0.30 originally). Shanks, Barbieri-Hermitte and Vadillo (2020) ran three studies of incidental price anchors and found no significant effect, while standard anchoring appeared robustly in the same work. Two preregistered replications of subliminal anchoring (Röseler et al., 2021) also failed.

What remains uncertain is the mechanism (insufficient adjustment, selective accessibility, or both) and how far the effect reaches beyond numbers people actively consider. The replication record points to a boundary: the anchor has to enter the judgment, not merely be nearby.