Why Loyalty Apps Use Random Rewards

Why We Come Back · 18 October 2026 · Reviewed by Faisal Zaidi

Every coffee app now spins a wheel, on the strength of a pigeon experiment from 1957 and a business book that was never a research claim. The one paper that tested repetition found the effect, and attached two conditions almost nobody copies.

Answer to the question

Do random rewards make people use an app again?

Yes, but only under two conditions. A 2019 paper led by Luxi Shen found across four experiments that people repeat a behaviour more when the incentive is uncertain than when it is certain, even when the certain one is worth more financially. The effect needs the uncertainty to resolve immediately, and appears only after a first repetition.

Key takeaways

The reinforcing-uncertainty effect, named by Luxi Shen and colleagues in 2019 across four experiments, holds only when the uncertainty resolves immediately and only after a first repetition.
In the 2015 motivating-uncertainty study, 70 per cent of people offered an uncertain reward worth less on average finished the task, against 43 per cent offered a certain reward.
A 2019 review and three experiments led by Jeffrey Pisklak failed to support the near-miss effect in both pigeons and humans, and questioned the standard paradigm.
A 2021 meta-analysis pooling 15 studies found a small to moderate correlation between loot box spending and problem gambling symptoms, with causal direction not established.

In four experiments run in 2015, people kept going for an uncertain prize more often than for a certain one worth more. Does that survive outside the laboratory, in an app someone opens next week? Random rewards are sold to retailers on the assumption that it does. The assumption has a famous source, a thin middle, and one real finding buried underneath.

A framework that cannot fail a replication

Where does the idea come from? From a model rather than a result. Nir Eyal's Hooked, first published in 2014 and updated in 2019, builds a four-part loop of trigger, action, variable reward and investment, and puts the unpredictable payoff at the centre of it. Eyal sorts those payoffs into rewards of the tribe, the hunt and the self, and argues that most gamification fails because the designer picked the wrong one.

We grade that book as a framework rather than as evidence, and the distinction is not a slight. A framework is an argued way of thinking. It cannot fail replication, because replication tests a measured claim and the Hook Model was never one. So the loop and the three reward types get cited as Eyal's taxonomy, with his name in the sentence, and never as proof that anything works.

The borrowed machinery is a different matter.

That part is solid, and it is much older than the book.

Schedules of Reinforcement, by Charles Ferster and B. F. Skinner in 1957, ran programme after programme on pigeons to see what different payout rules did to the rate of pecking. Rules that paid out unpredictably produced the fastest and steadiest responding, and the greatest resistance to extinction, meaning the behaviour carried on longest once the rewards stopped. Under an unpredictable rule, a reward that never arrives is indistinguishable from an ordinary gap.

Credit Ferster first, which almost no popular account does. And keep the scope: these are rates of key-pecking in pigeons under laboratory schedules. Nothing in the book concerns a human being deciding whether to buy a second coffee. The step from pigeon schedules to random rewards in a shop is the one Hooked takes without stopping.

Two different effects wearing a single name

If the pigeon work does not cover people and products, what does? Two separate literatures, and the trouble in this field is that both get called the same thing. One is about how hard you work inside a single pursuit. The other is about whether you come back. Random rewards are sold on the second and evidenced mostly on the first.

The first has a name: the motivating-uncertainty effect. In a 2015 paper led by Luxi Shen, then a doctoral student at Chicago Booth, 87 students were asked to drink water for a reward that was either a certain two dollars or an uncertain draw of one or two. The uncertain option was worth less on average. Seventy per cent of the people offered it finished the task, against 43 per cent of those offered the certain reward.

A follow-up with 530 people online varied the odds across eight bonus conditions and found something the design world rarely repeats: participants noticed that an outcome was uncertain and were largely insensitive to how uncertain. One per cent and ninety-nine per cent did similar work.

Then the authors bounded their own effect. In an auction for truffles, uncertainty raised what people paid while they were actually bidding, and lowered what they said they would pay in advance. Judging the offer in prospect produced ordinary uncertainty aversion, the plain dislike of a gamble that most of economics assumes. The comparisons rest on roughly 23 participants per cell, so this is the authors' own boundary rather than strong evidence, and it is the boundary that matters commercially: the customer reading your offer is judging, and the customer tapping the app is pursuing.

The condition almost nobody bothers to copy

Does any of that tell you whether people return? Not directly, because effort inside a task is not a repeat visit. The paper that asks the right question came in 2019, again with Shen as first author, in the Journal of Consumer Research, and it names its own finding the reinforcing-uncertainty effect. Four experiments, in the lab and in the field, with real consequences for every participant, small rewards and large.

The finding is awkward for economics and good news for loyalty managers. People repeat a behaviour more when its incentive is uncertain than when it is certain, in the authors' words "even when the certain incentive is financially better". That is the strongest published support random rewards have, and it is about repetition rather than about a one-off choice.

It arrives with two conditions, both stated by the authors, and the second is the one no loyalty programme seems to notice.

The uncertainty has to resolve immediately. If the resolution is delayed, the effect does not appear. A scratch panel that reveals itself at the till qualifies. A prize draw settled at the end of the month does not.
The effect shows up only after someone has already repeated the behaviour once, not before. It reinforces a customer you have. It does not recruit one you do not.

The reinforcing-uncertainty effect appears only after a first repetition. It reinforces a customer you have, and does not recruit one you do not.

The mechanism the authors favour explains both: the resolution is itself the reward, which is why it has to happen while the person is still there, and why they cast doubt on accounts built on reference points instead.

Read against our earlier piece on probabilistic price promotions, the distinction holds. A mystery discount at a single checkout and a random reward on the fifth visit are not one phenomenon with two names.

The near miss that did not hold up

So why do loyalty wheels look so much like slot machines? Because the design borrows a second idea, the near miss, and that one is in worse shape than the first.

The original finding is a 2009 paper in Neuron led by Luke Clark, then at Cambridge. On a simplified slot machine task, outcomes that fell just short were rated as less pleasant than clear losses and yet increased the desire to keep playing, while recruiting the same brain circuitry that responded to wins. The part that rarely travels is the control condition. The motivational effect appeared only on trials where the participant had arranged the gamble themselves. A wheel that spins on your behalf is not that condition, and most random rewards in a loyalty app spin on your behalf.

Ten years later, a review and three experiments led by Jeffrey Pisklak at the University of Alberta set out to pin the effect down and could not. A tightly controlled extinction procedure in pigeons failed to support it. The same procedure extended to human participants failed to support it. A third experiment, simplified further, found a clear conditional reinforcer in the reel and then no reinforcing function at all in extinction, which puts a question mark over the standard paradigm as well as the result.

Their conclusion is careful and we will keep it careful: there is no conclusive, replicable evidence that near misses sustain play, which is not the same as showing that no such effect exists. They called for replication, and for work with people who actually have a gambling problem.

What the loot box research actually shows

If random rewards do change repeat behaviour, who gets hurt? The nearest thing to an answer is the loot box, a paid in-game package whose contents are random. It is the purest commercial case of random rewards anyone has studied at scale, and the evidence there is consistent and correlational.

A 2018 survey of 7,422 gamers by David Zendle and Paul Cairns found that spending on loot boxes tracked problem gambling severity, and that the link was much stronger than the link between problem gambling and spending real money on other in-game items. The authors say plainly that the design cannot tell whether loot boxes are a gateway or whether heavy gamblers are drawn to them. A published correction the following year removed a stronger variance claim from the paper's conclusions, which the authors withdrew; the association itself stands. A 2019 replication on 1,172 gamers, built to fix the self-selected sample and the transparency of the first study's aims, reproduced a similar-sized association.

Zendle then looked at what happens when the mechanism is switched off. In 112 players of Heroes of the Storm, studied before and after the developer removed loot boxes, problem gamblers appeared to spend significantly less in-game afterwards, in contrast to other players. One game, one subgroup, no randomisation, and still the most direct evidence in the set.

Pooling the field, a 2021 meta-analysis led by S. S. Garea, which combines the results of many studies into one estimate, put 15 studies together and found a small to moderate positive correlation between loot box spending and problem gambling symptoms, rising when corrected for the tendency of null results to go unpublished. The authors state that causal direction is not established, and that figure comes from the accepted manuscript rather than the published version of record.

What the field evidence still cannot say

We looked for a field experiment that randomised loyalty rewards and kept a holdout group, and found none in the public literature. The reason is partly arithmetic. Across 25 large field experiments on advertising, Randall Lewis and Justin Rao found the typical range of plausible returns wider than 100 percentage points, and calculated that an informative test can need more than ten million person-weeks. The firms that could afford that keep the results. The vendor case studies circulating under brand names are company self-reports with no published method, and we are not citing them.

Nor did we find any independent attempt to replicate the motivating-uncertainty effect, in either direction. What exists is the authors' own boundary work, which is honest and narrow.

The trade-off, then, is unpleasant for anyone building this. The conditions that make random rewards work are the conditions that make them hard to defend: resolution on the spot, which is exactly what a slot machine supplies, and no effect at all on the first-time customer you were hoping to acquire. Keep the certain discount and you keep a promise your customer can read before deciding.

Take the mechanism and you take the family resemblance with it, along with a literature on who suffers that nobody has yet been able to make causal. So the choice is not between a clever mechanic and a dull one. It is between a promise your customer can read in advance and weigh against the alternatives, and a design that works best under the one condition you would rather not describe out loud to the person it is being used on, which is that the result arrives before anybody has time to think about it.

Common questions

Do random rewards make people use an app again?

The evidence says yes, under two conditions. A 2019 paper led by Luxi Shen found across four experiments that people repeat a behaviour more when the incentive is uncertain than when it is certain, even when the certain one is worth more. The effect requires the uncertainty to resolve immediately, and appears only after someone has repeated the behaviour once.

Is the psychology in Hooked proven?

Not in the way it is usually quoted. Nir Eyal's Hook Model and his three reward types are an argued framework rather than a tested claim, so they cannot pass or fail replication. The borrowed foundation is solid: Charles Ferster and B. F. Skinner showed in 1957 that unpredictable payout schedules produce the steadiest responding in pigeons.

Are loyalty apps using the same trick as slot machines?

They use the same ingredient, and the slot machine evidence is weaker than people assume. A 2019 review and three experiments led by Jeffrey Pisklak failed to support the near-miss effect in pigeons or humans. Meanwhile a 2021 meta-analysis of 15 studies found a small to moderate correlation between random-reward spending in games and problem gambling, with no causal direction established.

Why not just give a bigger certain discount?

Because a stated discount is read and compared, and a random reward is experienced. The 2015 truffle auction by Luxi Shen and colleagues found uncertainty raised bids during live bidding and lowered them in advance statements, on small samples. If your customer is deciding whether an offer is good, certainty wins. If they are already in the process, uncertainty can hold them.

Sources

Ferster, C. B., & Skinner, B. F. (1957). Schedules of Reinforcement. Appleton-Century-Crofts. doi:10.1037/10627-000. Bibliographic record read at OpenAlex; author order checked, Ferster first. The page-count figure in the vault's concept note was not confirmed at the published record and is not used.

Shen, L., Hsee, C. K., & Talloen, J. H. (2019). 'The Fun and Function of Uncertainty: Uncertain Incentives Reinforce Repetition Decisions.' Journal of Consumer Research, 46(1), 69-81. doi:10.1093/jcr/ucy062. Abstract read in full, reconstructed from the OpenAlex inverted index; paper closed access, so sample sizes and the field-study figures are not stated.

Shen, L., Fishbach, A., & Hsee, C. K. (2015). 'The Motivating-Uncertainty Effect: Uncertainty Increases Resource Investment in the Process of Reward Pursuit.' Journal of Consumer Research, 41(5), 1301-1315. Full text read, author's own copy.

Pisklak, J. M., Yong, J. J. H., & Spetch, M. L. (2019). 'The Near-Miss Effect in Slot Machines: A Review and Experimental Analysis Over Half a Century Later.' Journal of Gambling Studies, 36(2), 611-632. doi:10.1007/s10899-019-09891-8. Abstract read in full, reconstructed from the OpenAlex inverted index; open access. Sample sizes not stated in the abstract and not used.

Clark, L., Lawrence, A. J., Astley-Jones, F., & Gray, N. (2009). 'Gambling Near-Misses Enhance Motivation to Gamble and Recruit Win-Related Brain Circuitry.' Neuron, 61(3), 481-490. doi:10.1016/j.neuron.2008.12.031. Abstract read in full, reconstructed from the OpenAlex inverted index; full text not reached, so no sample size is stated.

Zendle, D., & Cairns, P. (2018). 'Video game loot boxes are linked to problem gambling: Results of a large-scale survey.' PLOS ONE, 13(11), e0206767. doi:10.1371/journal.pone.0206767. Correction 2019: doi:10.1371/journal.pone.0214167. Abstract and repository record read. The withdrawn variance claim is not used.

Zendle, D., & Cairns, P. (2019). 'Loot boxes are again linked to problem gambling: Results of a replication study.' PLOS ONE, 14(3), e0213194. doi:10.1371/journal.pone.0213194. Record read.

Zendle, D. (2019). 'Problem gamblers spend less money when loot boxes are removed from a game: a before and after study of Heroes of the Storm.' PeerJ, 7, e7700. doi:10.7717/peerj.7700. Abstract read in full, reconstructed from the OpenAlex inverted index.

Garea, S. S., Drummond, A., Sauer, J. D., Hall, L. C., & Williams, M. N. (2021). 'Meta-analysis of the relationship between problem gambling, excessive gaming and loot box spending.' International Gambling Studies, 21(3). doi:10.1080/14459795.2021.1914705. Abstract read at the accepted-manuscript preprint, figshare.utas.edu.au; figures not verified against the published version of record.

Lewis, R. A., & Rao, J. M. (2015). 'The Unfavorable Economics of Measuring the Returns to Advertising.' Quarterly Journal of Economics, 130(4), 1941-1973. doi:10.1093/qje/qjv023. Abstract read in full at the OpenAlex record.

Eyal, N., with Hoover, R. (2019). Hooked: How to Build Habit-Forming Products. Penguin Business, updated edition; first published 2014 by Portfolio/Penguin. ISBN 978-0-241-18483-7. Read through pages 1 to 246 in the vault; graded FRAMEWORK, cited as the author's argued model and not as evidence.