High-frequency compound actions should be clustered

What it is

Change the quantity, then add to cart: in real purchases those two taps almost always arrive as a pair. If the stepper sits mid-card and Add is pinned to a footer outside the card, every purchase crosses the whole card. Action clustering asks that controls repeatedly chained by the same task sit as a cluster the hand can finish without travelling, not split to opposite edges for visual symmetry.

The object is the action sequence, not “looks related”. Two buttons that share a theme but are never pressed in a row need not share a cell. Two controls that do not look like family, if the logs show them as a pair, should live together.

Why it happens

A practised sequence becomes a motor program: taps inside a small neighbourhood stay closer to automatic than a commute between two sites. Split the pair and every two steps insert a displacement; the program breaks, and working memory has to hold the half-state “I changed quantity, I have not added”. The larger the travel, the easier an unrelated entry intercepts the path — favourite, tapped on the way from stepper to button.

Copy and paste sit on adjacent keys for this reason, not for looks. Filter-then-apply, search-then-submit, crop-then-confirm on a screen pay the same commuting bill.

Studying it

From logs or recordings extract real paired sequences (rate at which B follows A inside a short window), then measure the on-screen distance between A and B. Move high-frequency pairs closer and compare sequence time, mid-path misses, and interrupted sequences.

Independent variables: distance between the pair, number of intervening controls, reach envelope of the hand in use. Dependent variables: time to finish the sequence, count of inserted unrelated actions, success rate of the pair.

KLM can estimate travel first, but acceptance has to use real sequences: the “happy path” in a mockup often fails to match the pairs in the log.

Where it stops holding

A destructive action that always follows Save still must not cluster next to Save — that is a different isolation problem. Adaptive toolbars that swap buttons by context dismantle a cluster the expert already learned, and that is a regression. A wide window that pins the primary button to the right edge is cheap for a mouse and may throw a thumb out of the hot zone; cluster against the actual pointer. One-shot wizard steps can follow temporal order without being packed into a spatial cluster.

Applying it

• From a week of logs, list the top pairs where B follows A, and inspect each pair’s layout distance. Pull high-frequency pairs into one block.
• On a product card, put quantity and add on the same action strip; do not leave one in the body and one in a page footer.
• Do not wedge a carousel or a promo between a search field and its submit, or between filters and Apply.
• How to check: walk the top three logged pairs on the page and, with thumb or mouse, see whether the travel crosses an unrelated control. If it does, the cluster is not built.

Related

• Same group: F1.03.1 Spatial distance between related elements is cognitive distance · F1.03.2 Every saccade carries a re-acquisition cost
• Nearby: F1.04 Thumb zone and primary-action placement · F1.06 Positional stability of frequent actions
• Search terms: action clustering · sequential co-location · KLM