What 949 coffee cards say about why starting is the hard part

In 1934, Clark Hull described rats in a maze: they "run faster as they near the food box than at the beginning of the path" (p. 39). Seventy-odd years later, three business school researchers took the idea to a café on the campus where two of them taught. What they found is a shape that would explain why the first step drags, one randomised test of a fix, and a twist about what happens right after you finish.

The café, and 949 finished cards

The café ran a buy-ten-get-one-free stamp card. The researchers collected 949 completed cards, roughly 10,000 purchases (p. 42), and timed the gap between each coffee and the next.

The gaps got shorter. From a customer's first gap to their last the average fell by .7 days, "an average acceleration of 20% from the first to the last interpurchase time" (p. 43), and a day-level model over 29,076 customer-days found the same (p. 44; Table 1, p. 45). All of that was watched, not arranged. Exactly one contrast randomised real behaviour, meaning customers allocated by chance and then watched buying coffee: the experiment below. Two later studies randomised travellers too, but into hypothetical scenarios rather than anything they did.

What that curve is not

It is not proof. Nobody was assigned to go cardless, so this is a pattern within people over time, not an effect against a comparison. The nearest control, forty-two customers paid $20 to carry a non-redeemable card, slowed down instead, and they opted in (pp. 43, 46). The curve is also drawn only from cards that got finished: on the incomplete cards the researchers bought back there was no linear acceleration (β̂1 = .02, p > .1), and across all their observed days those customers slowed down (β̂1 = .3, p < .01) (pp. 46 to 47).

The raw curve is bumpy rather than a slide: it rises at stamp 5, bottoms at 7, climbs again at 8 and 9, and that first-to-last contrast is its only statistical claim (t = 2.6, p < .05) (Figure 3, p. 43). The tidy version is modelled. And when the model let acceleration vary between segments, only the largest, 58 percent of members, accelerated significantly; the other 42 percent did not, with the 29 percent segment missing significance at p = .13 (Table 2, p. 47). The paper's verb for what that segmentation shows is "inconsistent with" a hidden-differences account; it never says ruled out.

Two stamps nobody earned

Research assistants posing as café staff handed 108 customers signing up one of two cards, at random: ten empty boxes, or twelve boxes with two already stamped as a joining gift. Ten purchases stood between either card and the free coffee (p. 48).

The head-start group finished in 12.7 days against 15.6, "nearly three days or 20% faster" (t = 2.0, p < .05) (p. 48). That is about speed and nothing wider: it compares customers who finished, and with no per-arm sample sizes or completion rates the paper never says how many finished at all.

Why it works, in the paper's own words: the head-start group "started with a lower proportion of original distance remaining to the goal than did the control group (i.e., d_t+2 = .83 and d_t = 1.0, respectively)" (p. 48). Two boxes of twelve pre-filled move you from 100 percent of the journey left to 83 percent while the work stays put. They call it illusionary goal progress.

The comedown

And then the twist. Among 110 members who finished one card and started a second, the first two gaps on the new card were 3.1 and 2.7 days against 2.2 and 2.1 at the end of the old one, model-estimated figures, and the difference holds at p < .01 (p. 45). That is their card- one opening pace of 3.2 and 2.8 days, near enough: no faster person was left behind. These are the same people watched before and after, not a group assigned to skip the reward, and on the second card the gradient appeared again from 100 percent remaining.

The shape repeats away from coffee. In an online music-rating programme, people were likelier to end a session the further they had to go (β̂1 = .49, p < .01) (Table 5, p. 53), and 80 of 473 session endings, 17 percent, came at the farthest point from the goal, which is either just after a reward or at the start of a run at one, against 1 of 473, 0.2 percent, with one song to go (p. 53). Those are raw counts over unequal exposure: everyone passes through the first state, almost nobody the second.

What it suggests trying

Two moves, neither is a promise to work, but logical and costless to try. Do not leave a list at zero: write down the step you have already taken and tick it. Then pick the next step before you stand up.

Where Yappie fits

We make Yappie, a focus app with a pet beside the timer, so read this as an interested party. We build a first step small enough to be done by the time you have read it, so a session never opens at 100 percent of the distance left.

References

1. Kivetz, R., Urminsky, O., & Zheng, Y. (2006). The goal-gradient hypothesis resurrected: Purchase acceleration, illusionary goal progress, and customer retention. Journal of Marketing Research, 43(1), 39 to 58. https://doi.org/10.1509/jmkr.43.1.39↩