behavioral ux
Back to library

Goal gradient effect

The goal gradient effect is the tendency to invest more effort per step as the remaining distance to a goal shrinks. What drives it is not absolute distance but the share of the original distance still left.

Publish your listing
Final checks before it goes live
Progress80% complete
  • Photos
  • Pricing
  • House rules
  • Payout details
Without the goal gradient
Publish your listing
Final checks before it goes live
Progress2 steps left
  • Photos
  • Pricing
  • House rules
  • Payout details
With the goal gradient
Live demo

The hypothesis comes from behaviourism: Hull's “The goal gradient hypothesis and maze learning” (1932) proposed that the tendency to approach a reward grows with proximity to it, a claim Bonezzi, Brendl and De Angelis summarise as motivation increasing monotonically with closeness to the desired end state. Kivetz, Urminsky and Zheng model effort as a function of the proportion of the original distance still remaining, written as (r − n) / r where r is the total requirement and n is what you have banked. That denominator is the interesting part: it means motivation can be changed by moving the perceived starting line without changing a single real step, which is why the same ten purchases feel closer on a twelve-step card that starts with two stamps than on a ten-step card that starts empty.

The corollary is that the gradient is local to one goal, not a general speed-up. When the counter resets, effort drops back to the starting rate. That is the part of the mechanism most interfaces forget: a product that chains goals endlessly keeps paying the reset cost.

The canonical source is Kivetz, Urminsky and Zheng (2006). In a university café programme, they collected 949 redeemed ten-stamp cards covering roughly 10,000 coffee purchases (29,076 customer-days). Raw interpurchase times fell by 0.7 days from the first to the last stamp, an average acceleration of 20% (t = 2.6, p < .05); cards completed in 24.6 days on average against 29.4 days at the first observed rate, about five days or 16% faster. A discrete-time proportional hazard model with time-varying controls gave a negative linear goal-distance parameter and a negative quadratic one (b1 = −.03, b2 = −.04, both p < .01), so the speed-up was real but decelerating, and the model beat a flat one (chi-square = 18.8, df = 2, p < .01).

Three controls in the same paper do the real work. An experimental control group of 42 customers carried transparent cards that looked identical but could not be redeemed; those customers decelerated as they neared the tenth stamp (b1 = .3, p < .01), which rules out a sales time trend or coffee habituation. For 110 members who completed two cards in sequence, the first two interpurchase times on the second card (3.1 and 2.7 days) were slower than the last two on the first (2.2 and 2.1 days, p < .01) — post-reward resetting, the signature of a goal-local effect rather than a habit. And the illusionary-progress field experiment randomly gave 108 customers either a ten-stamp card or a twelve-stamp card with two bonus stamps already applied. Both groups had to buy ten coffees. The head-start group finished in 12.7 days against 15.6 days, nearly three days or 20% faster (t = 2.0, p < .05; medians 15 versus 10, Z = 2.1).

The same paper replicates the gradient in a digital task. In an internet music-rating programme, 148 members rated 14,866 songs across 472 visits to earn $25 certificates. Visits came faster as the 51-song goal approached (b1 = −.25, b2 = −.22, p < .01), ratings per visit rose (Tobit b1 = −13.0, b2 = −7.5, p < .01), and people were least likely to quit mid-session when close: of 473 observed session terminations, 17% (80 of 473) happened immediately after a certificate was earned and the counter reset, 8.7% with one song banked, and 0.2% with fifty songs banked.

Converging field evidence exists outside loyalty programmes. List and Lucking-Reiley (2002) mailed 3,000 solicitations in six randomised arms of 500 for a $3,000 capital campaign, announcing that 10%, 33% or 67% of the goal was already raised. Participation went from 3.4% to 8.4% and the mean gift from $11.88 to $35.36 across the no-refund arms; total contributions rose 560% from the 10% to the 67% condition. Kuppuswamy and Bayus (2017) report that in large-scale Kickstarter data, project support increases as a campaign nears its funding target and drops sharply once the target is reached — the same gradient plus the same reset.

When it holds

  • It holds for people who are still in the goal. The headline café gradient was estimated on redeemed cards only. The 73 incomplete cards Kivetz and colleagues bought back showed no linear acceleration at all (b1 = .02, p > .1) and deceleration once post-purchase days were included (b1 = .3, p < .01); the 36 incomplete rating certificates likewise decelerated (b1 = .7, p < .1), with a significant completion-by-acceleration interaction (bINT = −.5, Wald chi-square = 16.2, p < .01). The authors read this as defectors lacking goal motivation, and they did model unobserved heterogeneity in the acceleration parameter (58% of the sample in a significantly accelerating segment, 29% with the same coefficient at p = .13). But the survivorship point is structural: people who quit early never enter the near-goal zone, so the observed curve is drawn from the people the goal already worked on. The paper's note that detailed analyses ruling out the self-selection account are available on request means that control is not in the published text.
  • The economic size can be near zero for the customers who matter most. Hartmann and Viard (2008) fitted a dynamic structural model to a golf course's buy-ten-get-one-free programme (531 golfers, 193,815 observations). Frequent customers' purchase incentives were practically invariant to the number of credits, and that group — the top quartile by frequency — accounted for more than two-thirds of all purchases and over 80% of all rewards earned. Infrequent customers did face substantial switching costs near a reward, but most left the programme before reaching half the required credits.
  • The shape is not reliably monotonic. Bonezzi, Brendl and De Angelis (2011) found a dip in the middle: in a nine-round word game (N = 126), participants produced fewer solutions at round five (M = 7.13) than at round two (M = 8.39, t(373) = 3.36, p = .001, d = 0.35) or round eight (M = 8.45, t(373) = 3.51, p = .001, d = 0.36). With N = 137 and a $300 charity goal, the frame decided the direction (Frame × Progress F(2, 131) = 3.525, p = .032): told how much was still needed, people gave most near the end ($2.86 versus $1.10 in the middle); told how much was already collected, they gave most at the start ($2.68 versus $1.12). A proofreading study (N = 69) reproduced the U with no frame at all: 0.122 typos per second near the start, 0.092 in the middle, 0.124 near the end (both contrasts p < .001, d = 0.55 and 0.63). These are small samples, but they show the gradient is a property of how progress is monitored, not of the task. Koo and Fishbach (2012) make the same point as a rule: motivation follows attention to whichever area is smaller, so accumulated progress motivates early and remaining progress motivates late.
  • The interface instantiation has its own, weaker record. Villar, Callegaro and Yang (2013) meta-analysed 32 randomised experiments from 19 studies on web-survey progress indicators. A constant-speed indicator did not reduce drop-off; the pooled effect was slightly in the wrong direction (log odds ratio 0.072, p = .365) and significant in only 2 of the 18 experiments. Pacing decided the outcome: slow-to-fast indicators raised drop-off, with odds 1.56 times the no-indicator control (log odds ratio 0.447, p = .001), while fast-to-slow indicators reduced it once one outlier was removed (log odds ratio −0.212, p = .02, odds ratio 0.80). In studies with a small completion incentive, showing a constant indicator raised drop-off by a factor of 1.17 after removing two outliers (log odds ratio 0.157, p = .025). So the gradient is not a licence to add a progress bar. Showing progress only helps when the progress you show looks fast early.
  • Expect the effect to be absent or reversed when the reward is not contingent on the behaviour (the transparent-card group decelerated), when the counter has just reset, when the remaining distance is displayed as accumulated progress late in the task, and when the goal is far enough away that the first steps look slow.

How to apply

  • Count progress against the end state once people are past the halfway mark: “2 steps left”, not “80% done”. Early on, do the opposite and show what is banked. This is the small-area rule from Koo and Fishbach, and it is the only moderator in this literature with a clean interface translation.
  • Make the first steps genuinely fast. Front-load the cheap steps so the indicator moves quickly at the start, because the fast-to-slow pacing is the only indicator shape that reduced drop-off in the meta-analysis.
  • Credit real prior steps on the counter rather than padding it. A head start works through perceived proportional distance, so count the account you already created and the email you already verified.
  • Keep the total requirement fixed and visible. The gradient is computed against the original distance, so a requirement that grows mid-pursuit removes the thing driving the effect.
  • Carry the named patterns: progress bars with honest pacing, step counters that say how many steps remain, and onboarding checklists where every item is a real task with a visible end.
  • Plan for the reset. Post-reward resetting is a measured effect, not a hypothesis: after someone finishes, effort returns to the starting rate and session abandonment spikes. Give the next goal a warm start, or let the finish be a finish.
  • Keep remaining steps small in number rather than small in label. Nine steps described as “almost done” is not the same stimulus as two steps.
  • If people are abandoning near the end, look for a mid-task dip rather than a final-step problem. The stuck-in-the-middle findings put the trough around halfway.

What to avoid

  • Progress indicators that stall, crawl or jump backwards near the end. The strongest single number in this literature is that slow-to-fast pacing increased survey drop-off by a factor of 1.56.
  • Fabricated progress: pre-ticked steps the person never did, bars that advance without work, counters inflated with steps the product invented. Beyond the trust cost, in the EU a representation that deceives the average consumer about the extent of the trader's commitments or the main characteristics of what is offered, and is likely to change a transactional decision, is a misleading action under Article 6(1) of Directive 2005/29/EC. Fake progress and fake remaining-step counts sit squarely in that definition.
  • Resetting earned progress silently — clearing a half-finished form on validation error, expiring stamps without notice. You destroy the accumulated distance that the whole effect runs on.
  • Manufacturing goals that exist only to extend engagement. The ethical test is whether the finish line is one the person would choose: a goal that serves the user is fine, a goal that only serves session count is engagement extraction with a progress bar on it.
  • Treating the gradient as a population-wide lift. In the golf-programme analysis the customers who generate most of the volume were practically unaffected by their credit count, so an incentive built on near-goal acceleration can pay out mostly to people whose behaviour it never changed.
  • Chaining goals with no exit. If every completion immediately opens a new counter, you repeatedly place people at the point of maximum abandonment.

Real-world examples

Satispay points screen showing emerald level at 6,926 out of 10,000 with a filled progress bar and a plus 1000 marker.
Satispay states the tier as 6,926 of 10,000 on a filled bar, so the balance reads as a short stretch left rather than a number on its own.
Ulta Beauty rewards screen with a 2% ring towards platinum, $500 to reach it, $489.01 amount to go, and a points scale marked at 100, 200 and 250 points.
Ulta Beauty puts the distance to the next tier in money — $489.01 to go — instead of leaving it as a level name, so the remaining gap is something the shopper can act on.
TheFork Yums screen at 0 points with a line reading only 1000 Yums to go to get €20 savings, on a scale running to €50 at 2000 Yums.
TheFork frames the next reward as only 1,000 Yums to go on a scale that already shows the one after it, so closing one gap opens the next.
1 / 3

Measuring it

  • It worked: step-level completion rate rises monotonically across the last third of the flow, and median time between steps falls as remaining steps go down. For repeating goals, compare interaction rate in the final quarter of a cycle against the first quarter.
  • It backfired: drop-off at the first step after the total requirement becomes visible, a drop-off peak near the midpoint, and a spike in abandonment in the session right after a goal completes. Watch the post-reset rate specifically — a one-off completion lift with a worse post-reset rate is a net loss.
  • Instrument the goal distance, not just the funnel. Logging remaining steps at each event is what lets you fit the curve instead of guessing at its shape.

References

For agents

An agent-friendly guide to this principle lives at /agents/goal-gradient-effect.md.

Our MCP server and skills for all principles are coming soon.

Similar principles

Disclaimer: The theories here are just that — theories. They may not apply universally and should be considered in context. Use your own judgment and consult additional sources when applying these principles.