The Paradox of Choice

The replication problem, stated plainly. The jam study was published in 2000 in the Journal of Personality and Social Psychology, one of the most selective journals in the field. It is elegant, it was run in a real shop with real shoppers, and its numbers are dramatic. It is also, on its own, one field experiment. When Scheibehenne, Rainer Greifeneder and Peter Todd assembled every comparable experiment they could find in 2010 — published and unpublished, roughly fifty comparisons — the mean effect of assortment size on choice was indistinguishable from zero, with enormous variance between studies. Their conclusion was not that the jam result was fabricated or that the effect is imaginary. It was that no one could yet say when it appears, which is a devastating thing to be able to say about an idea that had by then been converted into retail strategy, public policy advice and several thousand management presentations.

Two structural features of research make this outcome likely rather than freakish. Journals prefer surprising positive results, so failed replications of a beloved finding tend not to get written up, and the published record drifts away from the true average — publication bias. And a study with few participants can only reach statistical significance if the effect it happens to measure is large, so small studies that do get published systematically overstate the size of what they found. A single striking result is, for both reasons, the least reliable kind of evidence there is, and the one most likely to end up in a book.

Being fair to Schwartz. He was reporting the best available evidence at the time he wrote, and the strong version of the claim was mainstream psychology in 2004, not a fringe position. He has since acknowledged the meta-analytic results and moderated the claim, arguing that the effect is real but conditional — which is roughly where Chernev's meta-analysis also lands, and is a more defensible position than the one the book's subtitle advertises. It is worth separating the book from its reception: The Paradox of Choice contains a good deal of carefully hedged discussion, and the culture extracted from it a single memorable slogan and the jam.

The part that held up: maximisers and satisficers. This is Schwartz's more durable contribution, and it relocates the whole problem from the menu to the person.

The distinction descends from Herbert Simon, who coined satisficing in 1956 to describe an agent that searches until it finds an option meeting a threshold and then stops, rather than one that surveys everything and optimises. Schwartz, with Andrew Ward, John Monterosso, Sonja Lyubomirsky, Katherine White and Darrin Lehman, built a questionnaire in 2002 to measure where a person sits on that spectrum — whether they habitually search for the best or accept the good enough.

The finding, replicated many times since, is that maximisers report lower satisfaction than satisficers: less happiness with their decisions, more regret, more upward counterfactual thinking, lower reported optimism and life satisfaction. And in the most quotable result, Iyengar, Rachael Wells and Schwartz followed graduating students through a job search in 2006 and found that maximisers obtained jobs with starting salaries around 20 per cent higher than satisficers — and felt worse about them. The paper is titled "Doing Better but Feeling Worse", which is the entire finding in four words.

Note carefully what this does to the original thesis. It says the problem is not that the shop has two hundred jars. It is that some people respond to two hundred jars by attempting to evaluate two hundred jars, and that this is a losing strategy: the marginal improvement from more searching is small, the effort is large, and the vivid awareness of the rejected alternatives is corrosive. Satisficers standing in the same shop are fine. The overload is in the strategy, not the assortment — which is a considerably more interesting claim than "more is less", and one that generalises well beyond shopping.

This work is not beyond challenge either. Later researchers, notably Nenkov and colleagues in 2008, argued that Schwartz's maximising scale conflates several distinct things — high standards, extensive search, and difficulty deciding — and that they behave differently: high standards on their own are not associated with unhappiness, while the search-and-difficulty components are. If that is right, the finding is not "wanting the best makes you miserable" but "searching exhaustively and struggling to commit makes you miserable", which is both more plausible and more actionable.

The disagreement that remains. No one now defends the strong claim, and no one seriously argues the effect never occurs. The live question is whether the moderators identified so far are a genuine theory or a post-hoc list. A model that says overload happens under complexity, uncertainty and time pressure is much better than nothing, but it has not yet been used to make a risky forward prediction — to specify in advance which of two untested assortments will produce the effect and by how much. Until it is, the honest status is: a real phenomenon, unreliably summoned, imperfectly understood.

1956. Herbert Simon introduced satisficing, arguing that real decision makers with limited time and attention do not optimise but search until a threshold is met. The idea that the act of deciding has a cost enters economics here, and everything in this entry descends from it.

1970. Alvin Toffler's Future Shock coined overchoice, predicting that industrial societies would generate more options than their citizens could process. It was speculation without data, but it named the worry a generation before anyone tested it.

2000. Sheena Iyengar and Mark Lepper published "When Choice is Demotivating: Can One Desire Too Much of a Good Thing?" in the Journal of Personality and Social Psychology, volume 79, pages 995–1006. Three studies; the first is the jam table. The paper is careful, its authors were appropriately cautious about its scope, and it became one of the most cited findings in consumer psychology.

2002. Schwartz and colleagues published the maximising scale in the same journal, establishing the maximiser–satisficer literature.

2004. The Paradox of Choice: Why More Is Less appeared from Ecco. Schwartz combined the choice-overload experiments with his own work on maximising, regret, adaptation and opportunity cost, and drew a conclusion about affluent societies: that the expansion of options past a certain point had stopped adding to well-being and started subtracting from it. The book sold extremely well.

2005. Schwartz's TED talk carried the idea to an audience of millions and fixed the jam study in popular memory as an established fact. For most people who know the phrase "the paradox of choice", this is where they got it.

2010. Scheibehenne, Greifeneder and Todd published "Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload" in the Journal of Consumer Research, volume 37, pages 409–425. Mean effect near zero; large unexplained heterogeneity; no identified conditions. The popular idea and the research literature separated at this point and have not rejoined.

2015. Chernev, Böckenholt and Goodman published a second meta-analysis in the Journal of Consumer Psychology confirming the absent main effect but identifying four reliable moderators. This is the current mainstream position: the effect exists, conditionally, and the conditions are partly known.

Since. The episode became a standard teaching case in the replication crisis that reshaped psychology through the 2010s — cited alongside ego depletion and social priming as an instance where an appealing, widely believed effect turned out to rest on much less than its fame implied.

The most valuable thing about this entry is not the finding. It is the shape of what happened to it.

A single well-designed experiment produced a dramatic result. The result was memorable, told a satisfying story about modern life, and was picked up by a good popular book. From there it entered business advice, journalism, policy discussion and general knowledge — and it completed that entire journey before anyone had established that the effect reliably occurred. When the checking finally happened, ten years later, the average across fifty experiments was approximately zero.

Three practical lessons come out of that, and they apply well beyond this case.

A single striking study is a hypothesis, not a fact. However elegant the design, however respectable the journal, one experiment tells you that something happened once in one setting. The relevant question is never "was this study good?" but "what does the accumulated evidence look like?" — and for a genuinely new finding, the honest answer is usually "there isn't any yet."

Direction is not magnitude. "More choice reduces purchases" is a claim about direction and can be true in some settings while being useless. What you need is the size of the effect and how much it varies. An average effect of 0.02 with large scatter and an average of 0.5 with tight scatter both sound like "the effect is real" in prose, and they mean entirely different things in practice. Popular retelling almost always drops the number, because the number is the part that does not fit in a sentence.

Meta-analysis is where findings go to be tested. Pooling every study, including the boring ones and the ones that found nothing, is the only way to see past publication bias — the tendency for surprising positive results to be written up and unsurprising null results to sit in a drawer. It is unglamorous work that almost never produces a headline, and it is what turned "more is less" into the far more useful "more is sometimes less, and here is when."

There is a fourth lesson, easy to miss, and it argues against over-correcting. A near-zero average effect does not mean nothing is there. It can also mean that a real effect is being cancelled by an equally real effect running the other way, which is precisely what Chernev's moderators suggest is happening here. The correct response to a failed replication is rarely "the phenomenon is fake"; it is "the phenomenon is conditional, and we have not yet identified the conditions." Choice overload survived that transition, in reduced and better specified form.

And the practical residue is genuinely useful. If you are designing a menu, a form or a shelf, the number of options matters far less than whether the person in front of it knows what they want, whether the options can be compared without effort, and whether one of them is obviously reasonable. Structure, defaults and comparability do more than counting.

Streaming catalogues and recommender systems. The commercial answer to large assortments has not been to shrink them but to hide them. A catalogue of tens of thousands of titles is presented as a ranked row of about six, personalised. This is the moderator theory implemented as an interface: the underlying set stays enormous, while the set actually presented is small, ordered, and constructed so that one option looks like the obvious one. Whether this makes people happier than a genuinely smaller catalogue would is an open question, and one that the companies with the data have little reason to publish.

Retirement and health-insurance menus. The clearest policy application, and the one with real money attached. Pension plans and public insurance exchanges face exactly the conditions the moderators identify: options that are complex, preferences that are unclear, no dominant choice, and an outcome decades away. The interventions that work — default enrolment, tiered menus that present a handful of curated options before the full list, standardised comparison formats — do not reduce the number of available choices. They reduce the difficulty of the decision, which the evidence suggests is the variable that actually matters.

Dating apps and hiring. Both put a person in front of a large, weakly differentiated set with unclear preferences and no dominant option, and both generate the complaints the theory predicts: difficulty committing, persistent suspicion that a better option is one more swipe away, lower satisfaction with the choice eventually made. The maximiser–satisficer framing fits these better than the assortment-size framing does. The problem is less the size of the pool than the fact that the format invites exhaustive search and makes the rejected alternatives permanently visible.

Product ranges and the shelf. Retailers periodically cut their assortments, sometimes citing the jam study directly. The results have been mixed enough to be instructive: some ranges have been trimmed and sales rose, others were trimmed and customers went elsewhere for the item that disappeared. Which happens depends on whether shoppers arrive knowing what they want — and in a supermarket, mostly they do, which is why the effect is hard to reproduce there.

AI assistants and the shape of the question. Systems that answer with a single recommendation rather than a ranked list are, in effect, betting that the cost of choosing exceeds the value of choosing. The maximiser literature suggests that bet is often right and that the resulting satisfaction may be higher — and also names the price, which is that the person never sees the alternatives and cannot tell whether the answer was good. That is the same trade the sweet shop makes when it puts six jars on the counter instead of two hundred, with the significant difference that the six jars are visible and the filtering is not.