In this guide
Counting Without Stopping Fulfilment
A dark store cannot close for a count, which rules out the method almost every other kind of physical verification is built on. Orders arrive continuously, pickers are moving through the same aisles the counter needs, and the promised delivery window does not pause for verification. Cycle counting is the only workable model: the store is divided into zones, each zone is counted while it is briefly closed to picking, and the whole site is covered over a rolling schedule rather than in one night. What that costs is a clean cut-off. Because different zones are counted at different moments, the result is never a single snapshot, and stock moving between zones during the cycle can be double-counted or missed. That is recovered by freezing a zone properly before counting it, by capturing every pick and every replenishment against the zone during the freeze, and by reconciling the movements at the boundary rather than assuming there were none.
Why a Full Count Does Not Fit
The counting model that works everywhere else fails here for structural reasons rather than for want of effort. A delivery promise measured in minutes means the site is committed to fulfilling orders continuously during trading hours, and there is no version of a full count that does not stop that. The promise is the product, so suspending it is not a scheduling inconvenience but a withdrawal from the market for the duration. Freeze windows that do not exist compound it. Conventional counting relies on a period when nothing moves, typically overnight or on a closed day, and a quick-commerce site frequently trades for most of the day, receives replenishment outside those hours, and has no genuinely still period at all. The window that theory requires is not there to be booked. What a closed store costs per hour is the arithmetic that settles the argument. A site out of action loses not only the orders it would have fulfilled but the orders that route to a competitor and the customers who discover the competitor is adequate, and that second cost persists after the count has finished. Against that, the cost of a rolling model is modest even where it is less clean.
Zone and Slot-Based Counting
The workable model divides the site into zones and counts them in turn while the rest of the operation continues. Counting a zone while picking runs elsewhere is the basic mechanism, and it works because a dark store's picking is dispersed across the site rather than concentrated, so removing one zone from availability for a short period affects a fraction of the order lines rather than all of them. The zone is genuinely closed to picking for that period rather than nominally, since a zone being counted and picked simultaneously produces a result nobody can use. Locking a slot for the duration of a count is the finer-grained version, used where zones are too large to remove from service. The system blocks the specific locations being counted, the picker routing avoids them, and any order line requiring a locked slot is either held or substituted. Sequencing around the demand curve is what makes the whole approach tolerable. Counting the zones holding the fastest lines during the quietest part of the day, and the slow-moving zones during peak, keeps the interference with fulfilment at its minimum, and it requires the count schedule to be built from the order data rather than from the floor plan.
Handling Movement During the Count
Movement is the thing that breaks a rolling count, and the only workable approach is to capture it rather than to pretend it did not happen. Picks against a counted slot are the first case. Where a slot has been counted and an order subsequently draws from it before the cycle closes, the recorded count and the system position diverge legitimately, and the difference is a reconciling item rather than a variance. Capturing the picks against the slot during the window, with their timestamps, is what allows the reconciliation to be performed rather than argued about. Replenishment mid-count is the mirror case and is more disruptive, because inbound stock arriving into a location being counted can be counted, missed, or counted twice depending on timing. The practical rule is that replenishment into a locked slot is suspended for the duration, and where that is impossible the receipts are logged separately. Reconciling the movement rather than ignoring it is the discipline that distinguishes a rolling count that produces evidence from one that produces activity. Cycles that treat every difference as a variance, without adjusting for known movement, generate noise that eventually causes the whole programme to be disbelieved.
Evidence a Rolling Count Produces
A rolling count produces a different kind of evidence from a single-date count, and stating it as though it were the same overstates what it establishes. Coverage is expressed over the cycle: what proportion of locations and of value was verified across the period, and how recently each zone was last reached. There is no single date on which the whole site was verified, so a claim that the site was counted on a particular day is not available and should not be implied. Accuracy is reported by slot and by SKU class rather than as a site figure, because the whole purpose of the model is to show where the inaccuracy concentrates, and a blended number conceals exactly that. For an auditor to rely on the output, three conditions have to hold: the zone was genuinely frozen while it was counted, movements at the zone boundary during the freeze were captured and reconciled, and differences found were investigated and posted rather than adjusted. Rolling counts whose differences are written off without investigation provide activity rather than assurance, however complete the coverage looks.
Designing a Cycle for Your Store Network
Set count frequency against SKU velocity rather than counting everything equally. Fast-moving lines accumulate error quickly because they are touched constantly, so they justify weekly attention; slow-moving lines can be reached far less often without any real loss of control. A cycle that treats every SKU the same spends most of its effort where the least is happening, which is the commonest design fault in these programmes. Staff the count without pulling pickers, or accept that the count will be done badly. A picker asked to count between orders will prioritise the order, and the resulting sheet reflects that. Either bring in dedicated counters, or schedule the count into the genuinely quiet window the site has, which in most quick-commerce operations is narrow but real. An independent count is required anyway where a lender has charge over the stock, where the network figure supports a reported shrinkage number, or where the internal cycle has never been validated against an external one. Dark Stores & Quick Commerce work is planned around the operating clock rather than against it.
