Serial position effect

The curve is named after Murdock's 1962 free-recall experiments, and its two halves have different causes. Howard and Kahana state the asymmetry plainly: “In free recall, the recency effect is almost completely eliminated by 15 s of a distractor task (Glanzer & Cunitz, 1966; Postman & Phillips, 1965),” while the primacy end, built out of rehearsal and encoding, survives the same delay. Recency is not a short-term buffer either: put a distractor between every item as well and the end-of-list advantage comes back, a long-term recency effect “on time scales ranging from tenths of seconds to weeks.” So the last thing a user saw is privileged only while it is still the last thing. One confirmation screen, one interstitial, one upsell between your closing sentence and the decision, and the item you placed last is now a middle item.

Free recall is recall: the list is gone and the person rebuilds it. On a screen the list is still there, so order shows up in where the eye goes and what gets picked. Joachims and colleagues eye-tracked search sessions and found that split — abstracts at ranks 1 and 2 were viewed almost equally often, yet “users click substantially more often on the first than on the second link,” and with the top two swapped “there is still a strong bias to click on link one even if the second abstract is more relevant.” People scan downward: before a click at rank three, abstract two had been viewed 82.6% of the time and abstract one 73.9%, while the abstract just below a click is viewed only about half the time. Degrading the ranking moved the average clicked rank from 2.66 to 4.03, so position is a prior, not a fixed habit. Build as though the row under a tapped item was never seen.

At field scale, position moves attention and not preference. Ursu analysed a randomised ranking experiment at Expedia, where consumers got either the site's own ordering or a random one. Under the random ranking, position is unrelated to quality: hotels in positions 1–5 were 6.17 percentage points more likely to be clicked than those from position 21 down, 6–10 were 2.70 points more likely, 11–15 1.38 and 16–20 0.79, over 1.2 million displayed-hotel observations. Conditional on a click, position had no effect on purchase: “rankings have no causal effect on consumers' purchases conditional on search.” Under Expedia's own ordering the effect looked larger on both clicks and conversion, which is the point — an unrandomised ordering absorbs the quality it was sorted by. There is no end-of-list bump anywhere in these curves; being last in a scrollable list is just being far down it.

The cleanest causal evidence for a first-position advantage comes from ballots, and it is conditional. Ho and Imai exploited California's alphabet lottery, which randomises candidate order, across elections from 1978 to 2002. In general elections the effect on major-party candidates was negligible — a median loss of 0.4 percentage points, against a median gain of 0.2 for third-party candidates — while nonpartisan races gave first position about two points, and primaries, where party labels carry no information, produced a median effect of roughly 1.6 points, enough that the authors estimate ballot order “might have changed the winner in as many as 12 percent of all primary races examined.” Order decides in proportion to how little else there is to go on, which is a warning about a table of near-identical plans rather than a licence.

Which end wins depends on the channel, and that is where the folk version breaks. Holbrook, Krosnick, Moore and Tourangeau meta-analysed 548 response-order experiments inside 149 Gallup telephone surveys: heard rather than read, options showed recency, not primacy. Significant recency appeared in 19.2% of experiments and significant primacy in 1.8%, with a mean shift of 2.19 percent — so roughly four in five experiments found nothing significant, and the average effect is small. It grew with comprehension difficulty, with response options that were complete sentences rather than single words, with how many questions had already been asked, and with lower education (an average shift of 2.89 among the least educated against 0.57 among the most). The mechanism transfers to an IVR menu or a voice assistant: while option one is still being heard, option two interrupts the thinking, so the last option read is the one in mind when the thinking starts.

Where the list is visible, geometry decides, and neither geometry gives you a start-and-end rule. Höhne and Lenzner eye-tracked a web survey with 84 respondents and found primacy in vertical rating scales (Cohen's d = .24) and almost nothing in horizontal ones (d = .09), neither statistically significant; the eye data were far louder than the answers, with the first three categories fixated 2.97 times against 1.52 for the last three (d = 1.02). Atalay, Bodur and Rasolofoarison eye-tracked choices from 3 × 3 shelf planograms: brands in the horizontal centre drew more and longer fixations and were chosen 45.3% of the time against 27.3% for non-central brands, an 18.0 point advantage, while being on the left or the right made no difference at all — and the result held with the display shifted off the centre of the screen and with physical products on a shelf. So a vertical list of options leans weakly to the top, and a row of three pricing cards is governed by its centre.

Grade the parts separately. That people recall the start and the end of a vanished list is textbook-solid; that an interface should therefore put important things first and last is not one effect but three incompatible ones — top for a vertical list you read, centre for a horizontal array you compare, last for anything spoken — and no study here produces an end-of-list advantage in a list the user can still see. The measured effects are single-digit percentage points, mostly on which option gets examined rather than which gets chosen. Use order where order is all the user has, keep nothing load-bearing in the middle of a long list, and read “put it last” as a claim about what comes last before the decision, not about what sits at the bottom of the page.