Choice Overload Is a Conditional Effect: Four Moderators That Decide Whether Cutting Options Works

TL;DR: "Too many choices hurt conversion" descends almost entirely from one 2000 field experiment with jam. The 2010 meta-analysis of fifty experiments found a mean effect indistinguishable from zero with enormous variance across studies, and the 2015 meta-analysis showed why: overload is real but conditional on four moderators, namely choice set complexity, decision task difficulty, preference uncertainty, and decision goal. Field studies of actual assortment cuts show that perceived variety tracks shelf space and the availability of preferred items, not raw item count. The operator question is therefore never "should we reduce options" but "does this specific decision have the four properties that produce overload," and the reliable lever is usually comparison cost rather than option count.

---

A Folk Theorem With One Citation Behind It

There is a short list of behavioral science findings that made the jump from journal to boardroom so completely that practitioners no longer experience them as findings at all. Loss aversion is on it. Anchoring is on it. The endowment effect is on it. And near the top, cited in more product review decks than any of the others, is the claim that offering people more options makes them less likely to choose. It appears in pricing workshops as a justification for three tiers instead of six. It appears in navigation redesigns as a justification for collapsing a category tree. It appears in onboarding flows as a justification for hiding half the configuration surface behind a "more options" disclosure. Nobody argues with it, which is the first thing that should make an analyst suspicious.

The claim is almost always attributed, when it is attributed at all, to a single study. Iyengar and Lepper's jam experiment, published in 2000, is one of the most cited pieces of consumer research of the last three decades, and it has a rare property: it is legible. Two displays, twenty-four flavors versus six, and an order-of-magnitude difference in purchase rate. It fits on a slide. It survives retelling. It requires no statistical literacy to feel true. That combination is what makes a result travel, and travel it did, into design orthodoxy, into product strategy, into the standard vocabulary of merchandising teams who have never read the paper.

The literature did not stop in 2000

What almost never travels with it is anything published afterward. The literature did not stop in 2000. It ran a decade of replications, many of which failed. It produced a large meta-analysis in 2010 that found no reliable main effect whatsoever. It produced a second, more careful meta-analysis in 2015 that recovered the effect by conditioning on moderators. And in parallel, entirely separate from the psychology literature, marketing researchers ran field studies of what actually happens when a retailer cuts a real assortment in a real store, and found something considerably more useful than either camp of the psychology debate.

The question is never whether to reduce options

The synthesis of that literature does not support the folk theorem, and it does not refute it either. It supports something more demanding: choice overload is a conditional effect. It appears reliably under specific, identifiable conditions and disappears, or reverses, outside them. This means the question a product or merchandising team should be asking is not whether to reduce options. That question has no general answer and never did. The question is whether this particular decision, on this particular surface, for this particular visitor, has the properties that generate overload. Those properties are enumerable. They can be scored. And in most of the cases where teams reach for an assortment cut, at least two of the four are absent, which is why so many assortment cuts produce no conversion gain and a measurable loss in long-tail revenue.

This essay does four things. It reconstructs what the jam study actually established and what it cannot support. It walks through the 2010 meta-analysis, where the interesting number is not the mean but the variance. It works through the four moderators identified in the 2015 reconciliation, translating each into a diagnostic question an operator can answer about their own surface using data they already have. And it maps the whole apparatus onto digital surfaces where the assortment is not products at all, but pricing tiers, plan selectors, onboarding paths, feature menus, and search result counts.

What the Jam Study Actually Showed

The experiment is worth stating precisely, because the precision is where the trouble starts. Iyengar and Lepper (2000), publishing in the Journal of Personality and Social Psychology, ran a field study at an upscale grocery store in Menlo Park, California. A tasting booth was set up near the entrance. On alternating hourly rotations, the booth displayed either twenty-four varieties of Wilkin and Sons jam or six varieties. Shoppers who stopped at the booth could taste as many as they liked and received a discount coupon redeemable against a jam purchase. The dependent measures were how many passersby stopped, and how many of those who stopped subsequently bought.

The extensive display attracted more passersby than the limited one. That part of the result is consistent and rarely disputed, and it is the part practitioners most often forget: variety is an attractor. It pulls attention. Whatever else the study shows, it does not show that large assortments are bad at generating traffic, and a retailer who cuts assortment on overload grounds is trading away a documented attraction benefit for a hypothetical conversion benefit.

The conversion result is the famous one. Among shoppers who stopped at the booth, roughly thirty percent of those who encountered the six-variety display went on to purchase, against roughly three percent of those who encountered the twenty-four-variety display. An order of magnitude, in the field, in a real store, with real money.

The same paper reported two laboratory studies alongside the field experiment, one on chocolate selection and one on an optional extra-credit essay task, both contrasting a set of roughly thirty options against a set of six. Both found that the larger set produced less satisfaction with the eventual choice, or less follow-through on the choice, or both. The paper is not a single experiment. It is a package, and the package is internally coherent.

What the package can support is a narrow claim: under a set of conditions resembling a jam tasting booth, expanding an assortment from six to twenty-four reduced purchase conversion among people who had already engaged with the display. What it cannot support is the general claim that assortment size and conversion are inversely related across categories, channels, choosers, and decision types. That generalization requires the effect to be robust across contexts, and the whole subsequent decade of research was a test of exactly that.

One store, small counts, and the wrong endpoint

Three limitations deserve to be stated plainly rather than buried. First, this is one store, one category, one afternoon protocol. The external validity argument for extending it to software pricing pages, streaming catalogs, or B2B feature matrices is not an argument at all. It is an assumption. Second, the purchase-rate contrast rests on a small absolute number of purchasers. A ratio of thirty percent to three percent is dramatic as a ratio and considerably less dramatic as a count of human beings who bought jam, which is exactly the configuration in which sampling variability is largest and replication is least assured. Third, and most importantly for anyone running a business, the study measures conversion among booth visitors and does not measure category revenue. The extensive display attracted more people. The conversion rate among those people was lower. The paper does not settle what happens to total jam revenue, which is the number a retailer actually manages.

The Meta-Analysis Where the Mean Is Not the Story

By the late 2000s the replication record was messy enough to require a formal accounting. Scheibehenne, Greifeneder and Todd published that accounting in the Journal of Consumer Research in 2010, under a title that telegraphs the finding: Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload. They assembled fifty experiments, deliberately including unpublished studies alongside published ones, and computed a pooled effect.

The mean effect size was approximately zero.

That single number, taken alone, would license the opposite folk theorem: choice overload is not real, ignore it. This would be a mistake of the same kind as the original overgeneralization, made in the other direction, and Scheibehenne and colleagues were careful not to make it. The informative feature of their result is not the mean. It is the variance around the mean.

Across the fifty experiments, some found strong overload effects of roughly the magnitude the jam study reported. Some found no effect. And some found the reverse, with larger assortments producing higher choice rates, higher satisfaction, or both. The distribution is not a tight cluster around zero, which would indicate a genuinely absent phenomenon. It is a wide spread that happens to average out. Those are radically different epistemic situations. A wide spread averaging to zero is the statistical signature of an effect that exists but is governed by something the meta-analysis has not yet modeled. It is the signature of an unidentified moderator.

The published record was skewed toward confirmations

The authors said as much. Their conclusion was not that choice overload does not exist but that no reliable main effect exists, and that progress required identifying the conditions under which the effect appears. They also reported their own difficulties reproducing the original jam result in subsequent work, and the inclusion of unpublished null findings in the meta-analysis was a deliberate correction for the fact that the published record had been skewed toward positive results. A field in which the confirmations get published and the nulls sit in file drawers will produce a literature that looks far more settled than the underlying evidence warrants, and choice overload was a textbook instance.

A wide spread averaging to zero is not the absence of an effect. It is the presence of a moderator nobody has coded.

The Reconciliation: Four Moderators

The moderator identification arrived five years later. Chernev, Böckenholt and Goodman published Choice overload: A conceptual review and meta-analysis in the Journal of Consumer Psychology in 2015. Their contribution was to stop asking whether assortment size affects choice and start asking what else has to be true for it to do so. Once four moderators were entered into the model, choice overload became reliably observable and substantial in magnitude. Without them, it averaged to nothing, exactly as the 2010 analysis had reported.

The four moderators are not equally weighted in practice and they are not independent of each other. What follows treats each in turn, first as the literature defines it, then as a question an operator can actually answer about their own surface with data they already collect.

That translation step is where most applications of this literature fall apart, because the constructs are stated in the vocabulary of experimental psychology and the surfaces are described in the vocabulary of product analytics.

Moderator One: Choice Set Complexity

Choice set complexity is a property of the assortment's structure rather than its size, and it is the moderator most often confused with size itself. Chernev and colleagues decompose it into several components. The presence of a dominant or obviously attractive option collapses the decision regardless of how many other options exist. Attribute alignability governs whether options can be compared on a shared ruler or must be compared feature by feature with no common scale. Variance in option quality determines whether the set contains genuinely bad options that can be discarded quickly. Complementarity determines whether the options are substitutes competing for one slot or components that can be combined.

Alignability does most of the work in digital contexts and deserves its own definition, because it is the difference between a selector that reads as easy at twelve options and one that reads as hard at three.

Three pricing tiers that differ only on numeric limits are fully alignable. A prospective buyer compares seats against seats, storage against storage, and the decision reduces to a threshold check against their own usage. Three pricing tiers where each carries a bespoke bundle of non-overlapping capabilities are non-alignable, and the buyer must construct a mental model of which capabilities matter to them before they can compare anything at all. The second surface is harder at three options than the first is at twelve. Any framework that counts options and ignores this is measuring the wrong variable.

The dominant-option component is the one with the cleanest operating implication. If a visitor arrives at a set and one option is visibly better for someone with their profile, set size becomes close to irrelevant, because the comparison never has to happen. This is what a good default, a good recommendation, or a well-placed "most popular" marker actually does.

Diagnostic question. If a representative visitor had thirty seconds and were asked which option is best for them, would they be able to answer, or would they first have to build a comparison? The measurable proxy is dwell time on the selector before the first interaction, and the share of sessions that invoke a compare tool or open multiple options in parallel. High comparison behavior at low option counts is diagnostic of a non-alignability problem, and cutting options will not fix it.

Moderator Two: Decision Task Difficulty

Task difficulty is the moderator with the most obvious operational levers and the least attention paid to it. It covers time pressure, presentation format, the number of attributes visible simultaneously, and accountability, meaning whether the chooser will have to justify the decision to somebody else.

Presentation format is the piece that translates most directly. The same twelve options presented as a simultaneous grid, a paginated list, a filtered result set, or a sequential wizard produce four different difficulty profiles. Simultaneous presentation maximizes comparison opportunity and comparison load at once. Sequential presentation reduces load and also reduces the chooser's confidence that they have seen the best option. Neither format is universally correct, and which one is correct depends on the other three moderators — the recurring structure of this entire literature.

Viewport is a difficulty variable that did not exist when most of this research was designed and now dominates it. A twelve-option grid on a desktop monitor is a simultaneous presentation. The same grid on a phone is a sequential presentation with a memory requirement attached, because the chooser must hold earlier options in working memory while scrolling to later ones. Mobile traffic converts a comparison problem into a recall problem, and recall problems degrade much faster with set size than side-by-side comparison problems do. A team seeing an overload signature on mobile and not on desktop is seeing task difficulty, not assortment size, and the fix is presentation, not deletion.

Accountability runs counter to intuition. A chooser who must justify the decision to a procurement committee, a manager, or a spouse experiences more difficulty at a given set size than one choosing for themselves, because a defensible choice requires legible reasons and legible reasons require explicit comparison.

Diagnostic question. Under what conditions is this decision actually being made? The proxies are viewport distribution, whether the surface interrupts another task, how many attributes are simultaneously rendered, and whether the purchase is expensed or approved by a third party. A surface with high mobile share, many visible attributes, and an accountable buyer is difficulty-loaded before assortment size enters the picture.

Moderator Three: Preference Uncertainty

Preference uncertainty is, in my judgment, the most operationally important of the four, because it is the one that varies within a single surface rather than across surfaces. Set complexity and task difficulty are mostly properties of the design. Preference uncertainty is a property of the person, and the same page serves people at both extremes of it simultaneously.

The construct covers whether the chooser has articulated preferences, whether they have domain expertise, and whether they have prior familiarity with the specific options. A first-time visitor to an unfamiliar category has no preference structure to apply, so every option must be evaluated on its merits, and evaluation cost scales with set size. A returning power user who arrives knowing the exact model number has a preference structure that reduces the set to one regardless of how large it is. For that user, additional options are not a cost at all; they are the entire reason the catalog has value, because the option they want is somewhere in the tail.

This is where the folk theorem does the most damage. An assortment cut made on overload grounds imposes a real cost on the low-uncertainty segment, whose preferred items are disproportionately in the tail, in exchange for a hypothetical benefit to the high-uncertainty segment, which could have been served by a better default at no cost to anyone. The trade is almost always structured backwards, and the reason it survives is that the cost lands on a diffuse set of individually small transactions while the benefit, if it materializes, shows up in an aggregate conversion metric that leadership already watches.

The measurable proxies here are unusually good. New versus returning is a crude but available split. Entry path is better: a visitor arriving on a specific query has a preference structure, and a visitor arriving on a category browse likely does not. Best of all is behavioral, namely whether the session invokes filters, whether the account has a prior purchase in the category, and whether the visitor has saved or configured anything. Most commerce and SaaS teams have all three signals already and use none of them to condition the selector's design.

Moderator Four: Decision Goal

The fourth moderator asks what the chooser is trying to accomplish. Chernev and colleagues distinguish effort minimization from optimization, and immediate consumption from consumption deferred to a later point.

The effort-minimization case is the one the folk theorem describes correctly. A chooser who wants to be done, who does not regard this decision as worth investment, and who will accept any adequate option, experiences additional options as pure cost. Every option added is a longer scan and no additional value, because the first adequate option would have satisfied. For this chooser, a large set is genuinely harmful and a small set, or a strong default, is genuinely better.

The optimization case inverts it. A chooser who wants the best option, and who regards the search as worthwhile, gains from a larger set because the expected quality of the best available option rises with set size. Their satisfaction may fall, but their objective outcome improves and their likelihood of choosing does not necessarily fall. Cutting the assortment for this chooser removes the thing they came for.

The immediate-versus-deferred distinction matters more in digital contexts than it first appears. A visitor choosing what to watch tonight is in immediate consumption mode, and set size bites. A visitor building a watchlist, a wishlist, or a shortlist for a purchase two weeks out is in deferred mode, and a large set is an asset because they are not paying the comparison cost now. Surfaces that conflate the two, forcing a browse-to-learn visitor through a choose-now interaction, manufacture overload where none existed.

Diagnostic question. Is the visitor here to choose now or to learn the category? The proxy is sessions-to-conversion and the share of sessions that end with no selection at all but are followed by a return visit. A surface where most conversions happen on the third or fourth session is a learning surface being measured as a choosing surface, and its high single-session abandonment rate is not an overload symptom. It is the shape of the job the visitor is doing.

Loading diagram...

What Happens When Somebody Actually Cuts an Assortment

The psychology literature argues about laboratory choice tasks and tasting booths. The marketing literature ran the field experiment that operators actually care about, which is what happens to a real category in a real store when a real assortment gets cut. The answer turns out to be more useful than either side of the psychology debate, and it is barely represented in practitioner discourse.

Broniarczyk, Hoyer and McAlister published the foundational result in the Journal of Marketing Research in 1998, under the title "Consumers' Perceptions of the Assortment Offered in a Grocery Category: The Impact of Item Reduction." They studied what happens to perceived variety and to sales when a grocery retailer removes low-selling items from a category. The finding was that a substantial share of low-selling items could be removed without reducing either perceived variety or category sales, subject to two conditions. The consumer's favorite item had to remain available. And the category's shelf space had to be held constant.

Perceived variety tracks shelf space, not item count

That result reframes the entire problem. Perceived assortment is not a function of item count. It is a function of the physical space the category occupies and whether the specific item the shopper came for is present. A shopper does not enumerate the shelf. They form a gestalt impression from the footprint, then look for their item. If the footprint is the same size and their item is there, they do not notice that forty percent of the facings are now different products, and their behavior does not change. If the footprint shrinks, or their item is gone, they notice immediately, and the loss is not recovered by anything else on the shelf.

Boatwright and Nunes extended this into an online grocery context in the Journal of Marketing in 2001, with "Reducing Assortment: An Attribute-Based Approach." They studied a large assortment reduction and found that sales increased in most categories after a substantial item cut. The critical qualifier is that the effect depended on how the reduction was structured across attribute levels. Cutting within an attribute level, thinning the number of options that share a given characteristic, behaved differently from cutting an entire attribute level and removing that characteristic from the assortment altogether. A cut that preserves the range of attribute levels while thinning within them is a different intervention from one that eliminates whole branches of the category, even when both remove the same item count.

Put the two field results together and the operating instruction is specific and quite different from the folk theorem. It is not "reduce choice." It is: preserve the preferred item, preserve the navigational footprint, preserve the range of attribute levels, and cut the tail within levels. That is a surgical instruction with three constraints attached, and it bears almost no resemblance to the flat "fewer options convert better" advice that gets derived from the same body of research.

Translating to Surfaces Where the Assortment Is Not Products

Most of the decisions where this literature gets invoked in software are not product assortments at all. They are pricing tiers, plan selectors, onboarding paths, feature menus, integration directories, template galleries, and search result sets. The four moderators transfer, but the transfer requires being explicit about what plays the role of shelf space, what plays the role of the favorite item, and who the chooser is.

Shelf space, in a digital surface, is navigational footprint: the screen real estate the category occupies, its prominence in the information architecture, the number of entry points into it. The favorite item is the specific option the visitor came for, which may be an item, a plan, or a configuration. And the crucial difference from a physical shelf is that a digital surface can present a different footprint to different visitors, which means the Broniarczyk constraint can be satisfied per-visitor rather than globally.

Pricing tiers. This is where the folk theorem is applied most aggressively and least appropriately. Tier counts are usually small, three to five, which puts them well below any plausible overload threshold on count alone. The difficulty on a pricing page is almost never option count and almost always non-alignability: tiers that differ on bespoke feature bundles rather than on a shared numeric scale. The visitor cannot compare because there is no common ruler, not because there are too many rulers. Adding a fourth tier to an alignable pricing page costs almost nothing. Making a three-tier page non-alignable costs a great deal.

Plan selectors and configuration flows. Here decision goal splits sharply. A visitor configuring a plan during an active purchase is in choose-now mode. A visitor exploring what is possible before an internal budget conversation is in learn mode, frequently accountable, and will return. Treating the second visitor's non-conversion as an overload symptom, and responding by cutting configuration options, removes precisely the information they came to gather.

Onboarding paths. Preference uncertainty is at its maximum by definition: the user has no experience with the product and cannot have articulated preferences over paths whose consequences they cannot evaluate. This is the surface where the overload conditions are most consistently satisfied, and it is the surface where a narrow default genuinely outperforms a menu. The correct move is not to delete paths but to select one on the user's behalf and make the others reachable, which satisfies both the low-uncertainty minority and the high-uncertainty majority.

Feature menus for existing users. The mirror image. Preference uncertainty is low, expertise is high, decision goal is usually retrieval rather than choice. Overload conditions are largely absent, and the standard product instinct to simplify the menu for the sake of new users imposes a retrieval cost on the population that generates most of the usage. Search and command palettes exist because they resolve this tension without deletion.

Search result counts. Result count is the clearest case where the folk theorem is simply wrong. Nobody experiences overload from a result count of forty thousand, because nobody evaluates forty thousand results. They evaluate the first page, which is a set of ten. The relevant assortment is the consideration set the ranking produced, not the retrieval set, and the lever is ranking and filtering quality rather than index size. Reducing the index is not a simplification. It is a reduction in the probability that the visitor's preferred item exists at all, which is the exact constraint Broniarczyk and colleagues identified as the one that must not be violated.

The Lever Is Comparison Cost, Not Option Count

The practical consequence is that cutting options is the intervention with the worst risk profile in the entire available set. It is irreversible in practice, because reinstating a removed tail requires the operational machinery that was dismantled to remove it. It violates the Broniarczyk constraint by construction, since a cut deep enough to change the conversion rate is deep enough to remove some visitors preferred items. And it delivers its costs to the long tail, which is diffuse, individually small, and therefore invisible in the aggregate metric the cut is being evaluated on, while delivering its hypothetical benefit to a headline conversion rate that leadership is already watching. That asymmetry is why assortment cuts keep getting declared successes: the gain is measured and the loss is not.

The interventions that reduce comparison cost at constant option count carry none of this. Better default sort is reversible. A comparison view built on alignable attributes is reversible. A recommendation engine that surfaces a dominant option for a given visitor profile is reversible and can be held out for a control group. A filter taxonomy that lets a visitor collapse fifty thousand items into eleven is reversible. Each of these attacks the moderator that the 2015 meta-analysis identifies as first among the four, and none of them destroys tail revenue while doing so.

The exception, stated fairly, is the surface where all four moderators are simultaneously present and comparison-cost interventions have already been exhausted. Onboarding path selection for a genuinely novel product category is the clearest example. There, preference uncertainty is structurally maximal, no dominant option can be computed because nothing is known about the user, the decision goal is effort minimization because the user wants to start using the product rather than to study its configuration space, and task difficulty is elevated by the fact that the consequences of each path are opaque. That surface is a genuine overload candidate and should be narrowed. It is also, notably, a surface where the tail costs nothing to preserve behind a disclosure, which means even here the correct move is to narrow the default rather than to delete.

A Scoring Rubric for Deciding Whether Your Surface Qualifies

What follows is a diagnostic instrument, not a validated scale. The four dimensions come directly from Chernev and colleagues. The scoring bands and the thresholds are mine, constructed to force a team to state its assumptions in a form that can be checked against session data rather than asserted. They have not been validated against outcomes in any formal sense, and the cut points should be treated as conversation starters rather than decision rules.

Score each dimension from zero to three. Total ranges from zero to twelve.

The interpretation bands below are heuristic. They encode a deliberate asymmetry: because an assortment cut is irreversible and its costs are invisible in aggregate metrics, the burden of proof for cutting should be higher than the burden of proof for a comparison-cost intervention.

The scores are included because the rubric is easier to reason about when it is applied to cases readers already have intuitions about, and the pattern it produces matches those intuitions: onboarding menus for novel products and time-pressured checkout upsells score high, returning-shopper category listings and power-user command palettes score at or near zero.

Zero to four. Not an overload candidate. If the surface has a conversion problem, it is not an assortment-size problem, and cutting options will destroy tail revenue for no conversion gain. Look at ranking, defaults, page speed, pricing, or the offer itself.

Five to eight. Mixed. There is a real comparison-cost problem and it is very unlikely to be a count problem. Attack set complexity first: build an alignable comparison view, compute and surface a dominant option per visitor profile, improve the default sort. Re-score after those ship. In practice most surfaces that teams believe are overload candidates land in this band, and most of them resolve without touching the assortment.

Nine to twelve. Genuine overload candidate. Narrow the default set. Even here, the first move is to narrow what is presented by default rather than to remove options from the system, because the Broniarczyk constraint about preferred-item availability applies with full force and a disclosure costs nothing.

What This Changes on Monday

The immediate operational change is to stop treating "we have too many options" as a diagnosis. It is a hypothesis, and it is the least likely of the available hypotheses to be correct on any surface that has search, sort, filters, or a default. Before an assortment reduction reaches a roadmap, three things should be on the table: the rubric score with the evidence behind each dimension, the share of current revenue that sits in the items proposed for removal, and the specific comparison-cost intervention that was tried first and failed.

The second change is to instrument preference uncertainty. Almost every team has the raw signals, namely entry path, new versus returning, prior purchases or configuration, filter usage, and almost none of them condition the selector's design on those signals. A surface that presents a narrower default to visitors with category intent and no specific target, while leaving the full catalog one interaction away for everyone else, satisfies both segments and is testable with a clean holdout. That is, in the terms of the 1998 field study, the digital equivalent of holding shelf space constant while thinning the tail, the one intervention with field evidence behind it.

The third change is to the citation habit itself. The jam study is a fine study and it does not license the conclusion it is routinely used to license. When a research finding has been through two meta-analyses, the meta-analyses are the citation, and the correct summary of this one is that the effect exists conditionally, that four moderators determine when, and that the answer for any given surface is an empirical question about that surface rather than a general principle about human beings. That is a less quotable sentence than the jam story. It is also the one that survives contact with the evidence.

Key Takeaways

1. The popular claim that more options reduce conversion descends almost entirely from a single 2000 field experiment at one grocery store, in which purchase rates among booth visitors were roughly three percent for a twenty-four-variety display against roughly thirty percent for a six-variety display. That study also found the larger display attracted more traffic, a result that rarely travels with the conversion figure, and it measured choice rate conditional on engagement rather than category revenue.
2. Scheibehenne, Greifeneder and Todd (2010, Journal of Consumer Research) meta-analyzed fifty published and unpublished experiments and found a mean effect approximately equal to zero. The informative feature is the very large between-study variance, which is the statistical signature of a conditional effect governed by unmodeled moderators rather than of an absent phenomenon. The deliberate inclusion of unpublished nulls also indicates that the published record on this question had been skewed.
3. Chernev, Böckenholt and Goodman (2015, Journal of Consumer Psychology) recovered a substantial and reliable effect by conditioning on four moderators: choice set complexity, decision task difficulty, preference uncertainty, and decision goal. Each translates into a diagnostic question answerable from session data most teams already collect, and a surface that fails to satisfy at least three of the four is not an overload candidate.
4. The field evidence on real assortment cuts, chiefly Broniarczyk, Hoyer and McAlister (1998, Journal of Marketing Research) and Boatwright and Nunes (2001, Journal of Marketing), points somewhere more specific than "reduce choice." Perceived variety tracks shelf space and the availability of the shopper's preferred item rather than raw item count, and how a cut is structured across attribute levels matters more than how deep it is. The digital translation is to preserve the navigational footprint and the preferred item while thinning the tail within attribute levels.
5. The reliable lever is comparison cost, not option count, and the two decouple almost entirely once sorting, filtering, defaults, recommendations, and search are available. Comparison-cost interventions are reversible and testable with holdouts; assortment cuts are effectively irreversible, deliver their costs to a diffuse long tail that aggregate conversion metrics do not surface, and should therefore carry a materially higher burden of proof. Any assortment change evaluated on surface conversion rate alone cannot distinguish a simplification from an amputation.