Nov 10, 2026
6 min read
3 views

The first piece in our series "A claim you have heard: did it hold up?", from the 2000 jam study to two meta-analyses, with the verdict rule written down before we read any result.
Glasp Research. Keisuke Watanabe, Glasp Inc.
Too much choice paralyses us. Offer shoppers 24 jams instead of 6 and fewer of them buy, and those who buy are less happy with what they got. This is choice overload, and the jam study is its best-known example.
Our question: across the experiments that tested it, does a larger set, on average, lower the chance that people choose, or their satisfaction?
We could not read the original: it is closed access, and the public copies we found either refused our request or were scanned images. Its abstract says people were "more likely to purchase gourmet jams or chocolates or to undertake optional class essay assignments" with 6 options than with 24 or 30, and "reported greater subsequent satisfaction".
The source is Iyengar and Lepper (2000), in the Journal of Personality and Social Psychology. As the abstract and the two meta-analyses describe it, in the first of three experiments, a tasting table in an upscale grocery store showed either 6 or 24 jams. In the other two, people chose from 6 or 30 essay topics for a class, or from 6 or 30 chocolates.
The figures below are as the two meta-analyses describe it. More shoppers stopped at the 24-jam table, but 30% of those who saw the 6-jam display bought a jar, against 3% at the 24-jam display. In the chocolate study, satisfaction was 6.28 against 5.46 on a 7-point scale, and 16 of 33 people took chocolates instead of money as payment, against 4 of 34. One later paper gives the 6-jam purchase rate as 12%, not 30%. Without the original we cannot say which base each figure uses.
The 2010 meta-analysis reports, from its own authors' unpublished work, a repeat of the jam study in an upscale German supermarket that "did not find any negative effects of choice overload". It had the largest sample in that analysis, 504 people, and an effect size of d = 0.02, where zero means no difference. Lab repeats with chocolates (80 people, d = -0.20) and jelly beans (66 people, d = -0.27) also failed. We know these studies only from the meta-analysis.
We found no registered or multi-site replication of the jam study. A preregistered experiment from 2022 let 501 people give a small bonus to one of 2 to 80 charities, or keep it. The number of charities made no difference (95% interval for the effect of each extra charity, on the log-odds scale, -0.007 to 0.008).
Scheibehenne, Greifeneder and Todd (2010) pooled 63 conditions from 50 experiments with 5,036 people, published and unpublished. The average effect was d = 0.02, with a 95% interval from -0.09 to 0.12: "virtually zero", in their words, with large differences between studies. People with prior preferences or expertise did better with more options. The authors found "no sufficient conditions" for overload, and signs that published studies favoured it.
Chernev, Bockenholt and Goodman (2015) pooled 99 observations from 53 published studies with 7,202 people. Without moderators, their average effect was also not significant (the paper prints t(20) = -.10 and p = .48; those two do not match each other exactly, but neither reading is near significance). Their point was that the average hides opposite effects. In their model, a larger set did worse when options were harder to compare, when the task was harder (time pressure, more attributes, having to justify the pick), when people were unsure what they wanted, and when they aimed to spend little effort picking one item. All four factors were significant, with p below .001.
So the two papers agree that the plain average is not significant, and disagree on what follows. The 2015 authors report that once their four factors are in the model the overall effect of a larger set is significant, and that it is also significant (b = .17, p below .001) when studies of choosing between whole assortments are left out. A 2014 re-analysis that corrects for publication bias split the 2010 studies by direction. For studies where more choice looked bad, the corrected estimate was "effectively zero" (d = -0.07), though its authors note the data still allow sizeable effects of either sign.
Our rule, fixed in advance: take the largest preregistered meta-analysis or, as we found none, the largest with a heterogeneity-aware or bias-corrected estimate; if the 95% interval of its unconditional pooled effect includes zero, the claim does not hold as stated. That is the 2015 analysis. It prints no interval, but its unmoderated effect is not significant, which means the interval includes zero. The 2010 interval also includes zero, and the bias-corrected estimate is near zero.
The claim "more choice makes people choose less and like it less" does not hold as stated. A narrower version holds under stated conditions in one meta-analysis: hard-to-compare options, a hard task, unclear preferences, and an effort-saving goal.
It does not say choice overload never happens. Some studies find it, and a working paper (first posted in 2022, revised in 2026) with 1,832 people found clear overload in a task its authors call "arguably somewhat artificial" and chose because they expected overload there. They argue that the usual tests miss real effects.
It does not give a number of options that is "too many". In neither meta-analysis did the size of the larger set predict the effect.
The list of conditions is not settled. It comes from one team's coding of earlier studies, and we found no preregistered test of it.
We committed the protocol and verdict rule to our own git repository on 10 November 2026, before opening any source; that is our record, not a public registry. One planned citation search could not run because the database's free quota was used up.
Scheibehenne, Greifeneder and Todd (2010), author copy:
https://scheibehenne.com/ScheibehenneGreifenederTodd2010.pdf
Chernev, Bockenholt and Goodman (2015), author copy:
https://chernev.com/wp-content/uploads/2017/02/ChoiceOverload_JCP_2015.pdf
Iyengar and Lepper (2000), abstract:
https://pubmed.ncbi.nlm.nih.gov/11138768/
Simonsohn, Nelson and Simmons (2014):
https://doi.org/10.1177/1745691614553988
Lindkvist and Luke (2022):
https://doi.org/10.3389/fpsyg.2022.800528
Dean, Ravindran and Stoye, working paper:
https://arxiv.org/abs/2212.03931
This article was researched and written by Claude (an AI model by Anthropic) in a Glasp research run. No Glasp user data was used. Checks: read in full text were the two meta-analyses (Scheibehenne et al. 2010, pages 409 to 419; Chernev et al. 2015, pages 333 to 341 and 344 to 354), the choice overload section of the 2014 re-analysis, the 2022 charity experiment and sections 1, 2 and 4 of the working paper; read only as abstract were Iyengar and Lepper (2000) and McShane and Bockenholt (2018), and the original study's figures come from the two meta-analyses' descriptions of it; a second pass by the same AI session searched its saved copy of each source again and checked every number in this text against it, which is not an independent review; a separate AI run that did not write the text then compared the numbers with the sources on 11 November 2026, without being able to read the original study either. No human has checked these figures.