The first number anchors every subsequent judgment
Aliases: anchoring and adjustment · insufficient adjustment · selective accessibility
What it is
The anchoring effect: the first number to enter a judgment pulls subsequent estimates toward itself, serving as their starting point — even when that number is visibly random (spun on a rigged wheel of fortune) and entirely unrelated to the question. Classic demonstration: ask whether the tallest redwood is more or less than 1,200 feet, and later estimates of its height shift toward whatever anchor was given. Two boundaries: this is not rational use of reference information — a random number carries no information yet still works, so anchoring is a property of the estimation process itself; and it requires no active comparison on the user's part — a number works the moment it enters view and participates in judgment, which is nothing like deliberately shopping against a benchmark.
Why it happens
Two mechanisms, corresponding to two kinds of anchor. Numeric anchors run on anchoring and adjustment: people treat the anchor as a starting point and adjust toward what seems roughly right, but adjustment is effortful and stops at the edge of the first plausible range — insufficient adjustment leaves the estimate near the anchor. Adjustment scales with cognitive resources: time pressure and cognitive load shorten it, while explicitly instructed consideration of the opposite direction lengthens it and weakens the effect. Semantic anchors run on selective accessibility: the anchor enters as a candidate answer, the judge first retrieves evidence consistent with "above/below the anchor" to answer the comparative question, and that evidence stays accessible, contaminating the subsequent absolute estimate. Semantic anchors also rewrite the retrieval direction — which memories surface, which options get considered — so they move qualitative judgments ("is this worth it?") and not just numbers. Most anchors met in an interface are semantic: strikethrough prices, averages, and suggested quantities carry meaning, injecting a retrieval direction rather than a mere starting value. This also explains why warning users in advance barely helps: evidence made accessible does not become inaccessible again just because its source is known.
Studying it
- Paradigm: two classic families. The random-anchor paradigm: participants first receive a number from a rigged wheel of fortune, then estimate an unrelated quantity (such as the percentage of African nations in the United Nations), comparing mean shifts between high- and low-anchor groups — Tversky and Kahneman's wheel experiments established the floor result that random, irrelevant numbers still work. The comparative-then-absolute paradigm: first answer "higher or lower than the anchor," then give an absolute estimate, typically with much larger effects. Field studies use naturally occurring anchors as between-group contrasts: listing prices on property valuations, sentencing demands on judicial decisions, opening bids on final prices.
- Variables: anchor magnitude, anchor source (randomly provided / semantically related / self-generated — self-generated anchors ride more on adjustment and are more sensitive to cognitive load and time pressure), and adjustment resources; outcomes are mean shifts in absolute estimates, willingness to pay, and the correlation between estimate and anchor.
- Methodological cautions: the lab paradigm is a single judgment in an unfamiliar domain, where participants fall for the anchor knowing it is random; interface anchors are repeatedly exposed, semantically relevant, and met with purchasing experience — a different timescale entirely. Shift magnitudes measured in one-shot experiments do not transfer to "users who see strikethrough prices three times a week"; repeated exposure may internalize (reference values settle into internal benchmarks) or desensitize (users learn to ignore), and only longitudinal data can tell which. And a robust effect is not a large one: effect sizes differ enormously across paradigms, so cross-study comparison must hold the paradigm fixed.
Where it stops holding
Anchoring is among the most robust effects in behavioral research, but its strength has clear moderators. For decision makers experienced with the numbers it weakens but does not vanish: with an internal benchmark available (routinely purchased categories, familiar price levels), the anchor's pull shrinks; yet real-estate listing prices anchored professional appraisers too, who afterwards insisted they had been unaffected — expertise buys partial protection, and in a category with no benchmark the expert is as exposed as the novice. Anchor strength also tracks how relevant the anchor is perceived to be: anchors explicitly marked as irrelevant or random shrink but do not vanish (the wheel experiments announced their randomness), and skeptical users who suspect a marketing ploy shift less; conversely, any anchor that looks like information is taken at face value. Evidence boundary: classic results come from single-shot, low-stakes, Western-sample laboratory judgments, and shifts shrink in real purchasing decisions; "warning in advance eliminates anchoring" has repeatedly failed, at best slightly reducing it.
Applying it
- Audit every "first number" the interface generates — estimated delivery time, rating baseline, minimum order quantity, default capacity. Before showing a number, ask: will this become the starting point for the user's later judgments?
- Prefer a range or empty state over a single-point estimate that cannot be kept: once "arrives in 30 minutes" is shown, 45 minutes reads as a breach however sound the original basis — anchored expectations are settled in real satisfaction.
- Never name a number first in research interviews and surveys: asking budget before showing options and after produce two different budget distributions; start price tests with open questions and keep anchored phrasings as a separate experimental condition.
- To validate: run a first-number audit — record the first numeric value users meet between entering the decision context and deciding, and whether it relates to the task; then ask them in testing "where does your expectation come from" — anyone who can only parrot an interface number, with no independent source, has been anchored.