In short: De Koster, Le-Duc and Roodbergen put order picking at around 55 percent of warehouse operating expense in their 2007 review, with travel the dominant share of a picker's time, which is what makes warehouse slotting worth the effort. Velocity ranking treats each pick as an independent event and cannot see the tour, so correlated assignment, meaning items that appear on the same orders slotted near each other, is the part it misses. Travel distance per line picked is derivable from WMS task records and location coordinates, and it isolates the part of picker productivity that slotting controls. Every move costs labour and disruption, so candidate moves ranked by annual saving and stopped where payback crosses a threshold beats executing the whole list.
Walk a pick module on a Tuesday morning and the history is visible. The item at waist height in the first bay was a hero line in 2023 and now moves eleven cases a week. The item doing four hundred lines a week sits on the bottom shelf at the far end of aisle nine, because that is where a location came free the month it was introduced. Nobody decided any of this. It accumulated, one new SKU at a time, into whatever slot the last discontinued item vacated.
The accumulation is expensive because travel is the largest single component of a picker's time. De Koster, Le-Duc and Roodbergen, reviewing the order picking literature in the European Journal of Operational Research in 2007, put order picking at around 55 percent of warehouse operating expense, with travel the dominant share of the picker's time inside it. Very little else in the building offers that size of labour lever in exchange for a pallet truck and a system transaction, which is why warehouse slotting gets rediscovered every few years and then quietly abandoned once somebody prices the move list.
Velocity ranking is the first cut and the weakest one
The default slotting rule is popularity. Rank items by picks per period, put the fast ones near the despatch end, done.
A better version has existed since 1963, when Heskett published the cube-per-order index: rank items by the storage cube they consume divided by their order frequency, and give the closest locations to the lowest ratios. Close space is scarce, so an item earns it by generating picks per unit of space. A bulky item picked often can be worth keeping further away when it would otherwise displace four fast small ones.
Hausman, Schwarz and Graves added the practical refinement in 1976 with class-based storage. Sort items into a small number of classes, assign zones to classes, and let items float within a zone. The travel gain over a full rank ordering is modest and the stability is much better, because a rank ordering reshuffles every time demand moves and a three-class assignment mostly does not.
Both rules share a blind spot. They treat each pick line as an independent event, when what a picker does is walk a tour. The cost of a location depends on which other locations sit on the same tour, and popularity has no way to see that.
Affinity is the thing velocity cannot see
Frazelle and Sharp described correlated assignment in 1989: slot items that appear on the same orders near each other, because the tour is the unit that costs money.
The computation is straightforward and most operations have never run it. Take a year of order lines, build a co-occurrence count for every item pair, and normalise it. Raw co-occurrence favours pairs where both items are fast, so divide by the count you would expect if the two were independent. Pairs with a high ratio and enough absolute volume to matter are the candidates. Then cluster them: six items that regularly travel together should share a bay even when their individual velocities would scatter them across the module.
The effect concentrates in identifiable places. A parts distributor has kits that get ordered as kits. A grocery wholesaler has convenience store customers whose orders look nearly identical week to week. A pharmaceutical distributor has protocol-driven combinations. In those businesses affinity is worth more than velocity. In a business with genuinely uncorrelated single-line orders it is worth nothing, so check the co-occurrence data before buying anything that promises it.
Affinity and velocity want different layouts, and you cannot serve both by alternating between two rankings. The way out is to stop ranking and score a proposed layout directly: replay last quarter's orders against it under the routing policy your pickers actually follow, and total the travel. Ratliff and Rosenthal showed in 1983 that picker routing in a single-block warehouse is a solvable case of the travelling salesman problem, so tour length is computable rather than approximated. Evaluate against the route your terminal issues. A layout tuned for optimal routing can perform worse under an S-shape traversal than the layout it replaced.
The golden zone and the argument you will have about it
The golden zone is the band between roughly mid-thigh and shoulder height where a pick needs no bend, no stretch and no ladder. Petersen, Siu and Heiser studied golden zone storage in the International Journal of Operations and Production Management in 2005 and found the benefit concentrates heavily. Allocating that band to the fastest items is where the gain sits, and spreading it evenly across the range gives most of it away.
A second claim then arrives on the same space. The revised NIOSH lifting equation, published by Waters, Putz-Anderson and Garg in Ergonomics in 1993, computes a recommended weight limit as a product of multipliers, two of which are vertical origin and horizontal distance from the body. A 15 kg case lifted from ankle height at arm's reach falls far outside the recommended limit. The same case lifted from waist height close to the body sits inside it. The equation is explicit that heavy items belong in the golden band whatever their velocity says.
These two claims conflict and the conflict resolves by policy, since one side is a labour cost and the other is an injury risk. The workable arrangement is a hard constraint with an optimisation inside it. Anything above a weight threshold may only be slotted in the golden band, and the remaining golden space is allocated by picks per unit of face. That gives you a defensible answer when a health and safety review asks why the 18 kg case is on the floor, and it hands the slotting engine a smaller problem.
Travel distance per line picked
Almost every warehouse measures lines per hour. Very few measure travel distance per line, which is the number that tells you whether slotting is working.
Lines per hour confounds everything. It moves when the order profile changes, when a new picker joins, when the mezzanine is congested, when replenishment runs late and pickers wait at the face. Travel per line isolates the part that slotting controls.
The data already exists. WMS task records carry a location identifier and a timestamp per pick, in sequence within a task. Location identifiers almost always encode aisle, bay and level, so deriving a coordinate for every location is a parsing exercise plus a rack drawing. From coordinates and sequence you get path length per task under your routing rule, and dividing by lines gives the metric. A quarter of task history is plenty for a stable baseline by zone.
Track the distribution and not only the average. The mean moves slowly and hides the interesting part, which is the tail: tasks that walk three times the median distance. Those tasks point at specific items in specific locations, and that list is your reslotting worklist. It tends to beat any ranked report a slotting module produces, because it comes from what your pickers did rather than from what a model predicts they will do.
Reslotting has a bill and the saving has to cover it
A slotting run produces a move list, and the move list carries a cost that rarely reaches the business case.
Each move is a pick, a travel leg, a putaway and at least one system transaction, so call it three to six minutes of labour for a full pallet and more for a partial. The location is unavailable while the move happens, which means either running outside picking hours at a premium or running during them and creating congestion in the aisles you are trying to speed up. There is also a cost nobody models: experienced pickers carry a mental map of the module, and a large reslot invalidates it for a week or two, during which rates dip and mispicks rise.
The arithmetic per move is easy. An item picked twenty times a week, moved to a location three seconds closer, saves about 52 minutes of walking a year. At a fully loaded 25 an hour that is roughly 22 a year against a move cost of a few units of labour, so it pays back inside a month. The same three second improvement on an item picked twice a week returns about 2 a year and never pays back.
That asymmetry is the entire design of a sane reslotting programme. Rank candidate moves by annual saving, work down the list until payback crosses your threshold, and stop there. Most of the value sits in a small fraction of the moves. The long tail is where slotting projects go wrong, because the team executes four thousand moves, absorbs the disruption of all four thousand, and collects the benefit of the first three hundred.
Seasonal reslotting, continuous reslotting, and the case picking difference
Most operations should run two modes at once. A seasonal reslot is a planned event ahead of a known demand shift: the pre-peak layout in a business with a real peak, or a range change in a business with seasons. It suits a stable core range with predictable timing and it needs a quiet window, which is the part that usually gets squeezed. Continuous reslotting generates a handful of moves each week from the same ranked list and executes them in low-volume hours. It suits ranges with steady churn and no quiet window, and it sidesteps the mental map problem, since ten moves a week is invisible to a picker and a thousand moves over a weekend is not.
The two picking modes want different things from all of this. Full case and pallet picking is dominated by travel and by build sequence, because the pallet has to be stacked in an order that survives the trailer, and what happens to it after the dock door is a separate problem (X6). Each picking is dominated by the forward pick allocation question: which items get a forward location at all, and how much face each one gets. That allocation is a knapsack. A bigger face means fewer replenishment trips and more consumed golden space, so the item with the highest picks per cube of face earns the slot, while an item with a low ratio is cheaper to pick from reserve on demand. Operations that skip this decision end up with a forward area holding one facing of everything, which spreads travel across the whole range and gives the fastest movers a replenishment trip every two hours.
Where this stops
Slotting optimisation is a commitment to a demand mix. Moves pay back over weeks and months, and that payback assumes the ranking which justified them still describes reality when the benefit is being collected. In a fast-changing range the assumption fails. Fashion, promotion-driven grocery, marketplace assortments with weekly additions: the item ranked twelfth in April is unranked by July, and the reslot that chased it has already been paid for.
The honest answer in those environments is to slot less precisely and more durably. Wide classes instead of rank ordering, reslotting confined to the extremes, and a layout that is decent under many demand mixes. That is close to the argument Hausman, Schwarz and Graves made for classes in the first place, and the move cost arithmetic points the same way.
Two other things will stop a slotting project before the method does. The first is master data. Item cube and weight are wrong in most item masters, sometimes by a factor, and a slotting engine fed bad dimensions will confidently propose locations the product does not fit into, which is how a project loses the dock's confidence in week one. Measure a sample of a hundred items before running anything. The second is layout. Where everything funnels through one aisle or one lift, congestion is the binding constraint and rearranging items inside a bottleneck releases very little. Slotting also has nothing to say about which site should hold the stock at all, which is a network question with its own literature (I1).
Ahead of any of that, get the baseline. Pull a quarter of WMS pick task records this week, map the location identifiers to coordinates, and compute travel distance per line by zone, because without that number you will not be able to tell afterwards whether the reslot worked.