Compromise effect

Simonson (1989) built the effect on reasons rather than utilities: you pick the option you can most easily defend, and the one that is not worst on anything is easiest to defend. Across five product categories an option's share was on average 17.5 percentage points higher as the middle of its set than as an extreme, significant in nine of eleven comparisons: a $299 television took 23% of choices as the cheap edge and 48% as the middle, a 12-hour calculator battery 40% against 66%. The justification story only half held: expecting to defend the choice out loud moved the compromise effect by 2.6 points, not significant, though the same manipulation did move the attraction effect. Build for a visible trade-off, not for people rehearsing a reason.

Simonson and Tversky (1992) named the mechanism extremeness aversion and ran it on real catalogue pages. Against a $169.99 Minolta X-370, a $239.99 Maxxum 3000i split subjects 50/50; adding a $469.99 Maxxum 7000i made it 22/57/21, lifting the 3000i's share against the cheap body from 50% to 72%. The shift held among subjects who had first reviewed five cameras, ruling out the third option merely teaching people what the market costs. A dearer tier changes how a price you already sell reads.

The effect survives careful replication. Kivetz, Netzer and Srinivasan (2004) ran about a thousand respondents through overlapping triples of Pareto-optimal portable PCs and speakers: all eight tests were significant, with relative-share shifts of 15% to 34%, mean 24%. A second study of five PCs held the direction in all five contrasts and significance in four, 2% to 32%, mean 17%, with price an attribute for speakers and absent for PCs. Neumann, Böckenholt and Sinha's (2016) meta-analysis of 142 experimental observations calls extremeness aversion robust while finding its magnitude varies up to threefold with method: price-quality trade-offs, nondurable categories and binary-to-trinary comparisons shrink it, whereas more trade-off dimensions, non-numeric attributes, high-quality extensions and utilitarian products enlarge it. Those moderators run opposite to the attraction effect's, which in the replication literature needs every attribute to be a bare number, so a feature table mixing words and numbers suits this effect and not a decoy.

Position is not the mechanism, and this is the finding most pricing-page advice gets wrong. Chernev (2004) showed the compromise is the attribute-balanced option — equal ratings across attributes — wherever it sits. With no balanced option present, adding an adjacent alternative raised the middle option's relative share by 6.1 percentage points, the textbook result. When the balanced option sat next to the middle instead, the middle option's relative share fell by 12.3 points, a significant reversal; when the balanced option was the middle one, it gained 9.3 points. Only 20.1% of respondents named the middle option as the compromise if an adjacent option was balanced, against 65.5% who named the balanced one. A middle tier that is plainly worst on a row is an edge, whatever the layout says — and this is where the effect parts from the decoy effect, which needs a dominated neighbour rather than a balanced middle.

Two conditions that every real interface violates weaken it. Pettibone (2012) gave people 2, 4, 6 or 8 seconds to choose: no significant preference for the compromise target at 2 or 4 seconds, a significant one at 6 and 8, a trend of about 6 percentage points per second of deliberation. Dhar and Simonson (2003) added the option not to choose anything and found it “strengthens the attraction effect, weakens the compromise effect, and decreases the relative share of an option that is ‘average’ on all dimensions,” because not choosing competes with the safe middle pick. A skimmed pricing table with an exit one click away is close to the worst case.

The field evidence is thin but real, and it points at the range rather than the middle slot. Ekström (2021) randomised 63,494 recipients of a mail fundraiser across three ask strings. Moving $100 from the top of ($10, $50, $100) into the middle of a wider ($10, $50, $100, $250, $500) almost tripled the share giving exactly $100 — a 0.041% base rising by 0.078 percentage points — and raised the donation rate by 0.21 points, so the narrow string had been turning donors away rather than trading them down. But a third string listing only the endpoints, $10 and $500 with a blank, performed identically on both measures (p = .89), with 90% of its donors writing in their own amount against 30% in the explicit-middle condition. Setting the range is the operative instruction; naming a middle is optional. The counts are small: 241 donors, 61 gifts of exactly $100. On soft drink sizes, Sharpe, Staelin and Huber (2008) found adding a 44-ounce option cut 32-ounce buyers' average purchase by 2.58 ounces; their demand model implies a 17% rise in consumption, 17.7 to 20.6 ounces per customer, for a chain swapping a 12-to-32-ounce line-up for a 16-to-44-ounce one — a policy simulation, not observed sales.

It fails in places that look like fair tests, and it is easy to confuse with the layout. Mckay, Kemps, Prichard and Tiggemann (2023) had 216 participants order from online menus carrying three or four of small, regular, large and extra-large: adding a small or an extra-large option did not significantly change how often the large was chosen, and most people took the size labelled “regular” whatever the range, which the authors read as a familiar label doing the work instead. A middle option can also win for reasons unrelated to trade-offs: Valenzuela and Raghubir (2009) found people infer that whatever sits in the centre of a display is the most popular, an effect that disappears when the layout is uninformative or their own preferences are made accessible, and Rodway, Schepman and Lambert (2012) replicated the pull of the centre of five items. If your middle card also carries the badge and the centre slot, you have not learned which of the three moved anything.

Context effects are not consumer-specific — Trueblood, Brown, Heathcote and Busemeyer (2013) produced all three in a perceptual-discrimination task — and the incentive-compatible numbers are the honest ones to plan around. Dumbalska, Li, Tsetsos and Summerfield (2020) found compromise the most reliable of the three classic decoys among 189 participants ranking rental properties for real stakes, with the similarity effect absent entirely, and each decoy shifted preference by about 3 to 5 percentage points. Huber, Payne and Puto (2014), defending their own attraction effect, concede that “although asymmetric dominance may be more robust than compromise in the lab, we more often observe successful use of compromise in the marketplace.” Size the work for a few points, not for the 17 to 24 point swings of forced-choice matrix studies, and not for any published pricing-page experiment, because there is none.