Von Restorff effect

Hedwig von Restorff's 1933 dissertation is in German; MacLeod's (2020) history went back to the source. Across nine experiments she gave people lists of paired associates built from a single material — nonsense syllables, numbers, letters, colours — with one or a few pairs isolated against that uniform background, then tested free recall. Isolates were recalled better than the “massed” items whatever the material. Two of her own observations are routinely dropped: recognition tests “yield smaller differences than do tests of reproduction,” and she read the effect as a cost to the crowded items, functionally disadvantaged by belonging to a subgroup, rather than as a bonus to the odd one. The original finding is about remembering a list later, not about what a user notices now.

Hunt's (1995) re-reading is the standard citation for what the effect requires, and it argues against the design slogan. Von Restorff herself presented evidence that perceptual salience is not necessary, and Hunt's point is that difference alone is not sufficient either: distinctiveness is relational, so an item is distinctive only against a background of similar items. Schmidt (1991) had already warned that the concept had been stretched across four classes of phenomena that behave differently, and concluded it was being overapplied. On a screen the work is done by the sameness of everything else, so a highlight dropped into an already-varied layout has nothing to contrast with.

The most careful recent test makes both the effect and its cost visible. Schmidt and Schmidt (2017) ran 15-word lists under free recall. An isolate mid-list was recalled .55 of the time against .43 for the same word in a homogeneous list — a marginal advantage — and against .27 when the whole list was unrelated, which is the uniform background doing the work. Meanwhile the non-isolated items around it fell from .57 to .53. Isolating an item in the second position did not help at all: isolates were recalled .44 against .63 in homogeneous lists, and the impairment replicated both when participants simply read the list and when they named colours while reading it. A cross-category contrast, a word among numbers, still gained early, but a merely conceptual oddity lost. Their conclusion is that “the early isolation effect may be of more limited generality than many researchers have assumed.” Emphasis near the top of a set can cost you the item you emphasised, and part of what the isolate gains comes off its neighbours.

Which kind of difference you use decides whether a distracted user gets anything. Bireta and Mazzei (2016) crossed full and divided attention with physically isolated items — a different colour or typeface — and semantically isolated ones. Under full attention both worked: physical isolates .69 against .47 for controls, semantic isolates .60 against .40. Under divided attention the physical effect survived (.51 against .31) while the semantic effect disappeared entirely (.34 against .35). A colour or weight change is close to automatic; a cleverly worded label is not, and a user glancing at a screen while doing something else is the divided-attention case.

None of that is about looking, and the looking literature is separate and better measured. In Wolfe's (2021) Guided Search model a target defined by one guiding feature such as colour gives a search slope “near (but typically a little greater than) 0 ms/item,” so a single unique element is found in about the same time however many others surround it, whereas serial search runs at roughly 90 to 350 ms per item. But guidance is not the same as salience: bottom-up salience is in Wolfe's account a powerful but short-lived form of guidance, and in real scenes meaning maps predict fixations better than salience maps do. Orquin, Lahm and Stojić (2021) meta-analysed 122 effect sizes from eye-movement studies of decision making and ranked the visual levers: central position ρ = .43, surface size ρ = .35, fewer competing elements ρ = .24, and visual salience last at ρ = .13. Where you put the element and how big you make it buy more looking than how loudly you colour it.

A distinct block on a working page can also be skipped precisely because it is distinct. Burke, Hornof, Nilsen and Gorman (2005) eye-tracked people searching news headlines: participants looked at banners in 11.7% of trials, and at plain grey placeholder rectangles (8.8%) almost as often as at designed static (13.0%) or animated (13.2%) ones. Recognition afterwards was at chance. The ads were not free: commercial banners slowed exact-match search against grey rectangles, flashing text raised perceived workload, and a target sandwiched between two banners took 75% — 4.3 seconds — longer to find. Pagendarm and Schaumburg's (2001) smaller study points the same way: aimless browsers remembered banners significantly better than people sent to find something. Shape a highlight like an ad and the user with a task will miss it and lose time to it.

The interface claim that has actually been measured is emphasis on a consent dialog, and it is worth single digits. Utz, Degeling, Fahl, Schaub and Holz (2019) ran notice variants past 82,890 visitors of an e-commerce site; in the confirmation-only notice, “Accept” drawn as a highlighted button was clicked by 50.8% of mobile visitors against 39.2% when the same word was a plain text link (26.9% against 21.1% on desktop). Structure did far more than styling: around 30% of mobile visitors accepted all third parties when the checkboxes came pre-ticked, against under 0.1% when they had to tick them, and Nouwens, Liccardi, Veale, Karger and Kagal (2020) found that notification style — a blocking barrier or a banner — did not affect the consent rate at all, while removing the “reject all” button from the first page raised acceptance by 22.0 and 23.1 percentage points. Emphasis can also run backwards: in Armel, Beaumel and Rangel's (2008) lab task, showing one of two real foods for longer raised its choice share by 6 to 11% for liked or neutral items and lowered it 7% for disliked ones. Nothing in the isolation literature measures choice, no published experiment tests a highlighted pricing tier, and the nearest measured analogue is a consent button worth about ten points on mobile — so treat the highlight as a scarce pointer at one option and measure the selection on your own screen.