Inbound receiving sets your opening balance, but accuracy is what keeps that balance true once thousands of small transactions have run through it. Overseas stock is also the hardest to verify: you cannot walk the floor, and a drift of a few units per SKU corrupts reorder points months before anyone notices. This guide covers how accuracy is measured, the counting methods that find errors while they are still explainable, and the clauses that decide who absorbs a shortage.
Inventory accuracy is not one metric among many — it is the quality of the data the others are built on. Reorder points fire on available stock, so a wrong balance orders too early or too late. Days-of-cover divides stock by velocity, so a wrong balance misstates how long stock lasts. Turnover and sell-through are calculated from the same counts; if they drift, every ratio built on them drifts too.
The reason is also commercial: an overstated record causes overselling, cancelled orders and marketplace penalties, while an understated one hides sellable stock you already paid to ship and store. Report the two directions separately rather than netting them into one figure.
A variance is any disagreement between record and physical reality, and quantity is only one dimension. Stock can be right in units and wrong in location, which makes it unpickable; wrong in status, where damaged or quarantined units sit in the available pool; wrong in lot or serial identity, which breaks traceability; or wrong in ownership. A discrepancy is a signal to investigate, not evidence of theft — the cause is usually procedural.
Before comparing two warehouses, or the same warehouse across two periods, confirm which definition is in use. Each answers a different question, and the same operation can report very different "accuracy" figures depending on the basis chosen.
| Basis | How it is calculated | What it tells you | Where it misleads |
|---|---|---|---|
| Record (SKU-location) accuracy | Records with no variance ÷ total records counted, within an agreed tolerance | How often the system is simply right; easiest to report and audit | A one-unit error on a high-value SKU counts the same as on a cheap one |
| Unit-based accuracy | 1 − (total absolute unit variance ÷ total system quantity counted) | The size of the error in units; uses absolute variance so shortages and overages do not cancel | Large low-value holdings can mask a serious shortfall on a small expensive line |
| Value-weighted accuracy | Variance valued at an agreed cost basis, measured against inventory value | Financial exposure, which is what a shrinkage claim turns on | Depends entirely on the agreed cost basis — standard, average or landed cost |
Quoted accuracy targets vary widely between sources, and much of the difference is definitional rather than real performance. Agree the basis, the tolerance and the counting population in writing first; only then is a trend meaningful.
Whether a one-unit difference counts as a failure is a policy decision, not a technical one. Three policies are common: zero tolerance, unit tolerance and a percentage share of on-hand quantity. Tolerances should be tighter on high-value or fast-moving lines, where a small variance cascades into missed orders, and looser on slow, low-value stock where investigating costs more than the exposure.
Keep the policy constant across reporting periods, and report positive and negative variances separately so an overage on one SKU cannot hide a shortage on another.
Cycle counting beats the annual stocktake for one practical reason: it surfaces a discrepancy while the documents that explain it still exist — a variance found eleven months later is almost never traceable. The methods differ in what they target, and most mature operations blend them.
| Method | How it works | Strength | Limitation |
|---|---|---|---|
| ABC / Pareto counting | SKUs tiered by value, velocity and risk; A counted weekly, B monthly, C quarterly | Concentrates effort where the financial and service risk is highest | Tiers go stale; reclassify when demand patterns shift |
| Random sample counting | Locations selected at random across the catalogue | Unbiased coverage of the long tail that ABC neglects | May miss high-risk SKUs if used alone |
| Control group counting | A small fixed set of locations counted repeatedly for several weeks | Tests whether the counting process itself is reliable before scaling it | Limited coverage; a diagnostic, not a programme |
| Opportunity-based counting | Count triggered by an event: empty bin, short pick, replenishment, discrepancy report | Almost no incremental labour; catches problems at the moment they appear | Irregular coverage — quiet zones can go months unchecked |
| Location or zone sweeps | One aisle, zone or bay counted end to end in sequence | Finds misplaced stock and location errors rather than quantity errors | Ignores item value; useful after re-slotting or a layout reset |
| Zero-bin counting | Locations the system believes are empty are confirmed empty | Flushes phantom stock in seconds; prevents a short pick later | Only useful as a routine habit, not a scheduled programme |
A schedule fails when it is ambitious rather than achievable. Work backwards: decide the period in which each tier must be fully covered, divide the SKU population by the working days available, and you have a daily count target. A modest volume completed daily beats a large one completed occasionally. Schedule counts in quiet windows — early morning before the first pick wave is usual — because counting a location while picks are in flight creates variances that are artefacts of timing, not real errors.
Assign an owner, publish the plan a shift ahead and freeze the count set. Freeze only the bins being counted, not whole zones: locking down large areas creates pressure to rush, and rushed counts are how counting errors enter the data.
Three controls separate a programme that improves accuracy from one that merely generates adjustment entries. Count blind: the counter receives the SKU and location but not the expected quantity, because a visible system figure turns a test into a confirmation. Separate duties: the person who counts should not approve the adjustment, and ideally should not normally pick or receive in that zone. And rotate counters between zones, since a counter who always counts the same shelf develops unconscious assumptions about what should be there.
Resist suspending counting during peak season: peak months carry the highest transaction volume and therefore the highest error rate. Reducing count volume is reasonable; stopping it is not.
Variances are overwhelmingly procedural. Theft and fraud are possible, but they should be a conclusion reached after evidence, not the first explanation offered. The recurring causes look like this:
| Cause | What it looks like | Where to look |
|---|---|---|
| Receiving errors | A wrong opening balance that every later transaction inherits | ASN versus blind count at the dock; short or duplicate receipts |
| Put-away and location errors | Stock physically present but the system points elsewhere | Unlabelled bins, dual locations, temporary overflow never recorded |
| Unrecorded movements | Stock moved physically, the system task abandoned | Transfers, replenishments, scrapped or damaged units pulled without posting |
| Unit-of-measure errors | Cases, inner packs and singles converted inconsistently | Pack configuration in the item master versus what the carton contains |
| Master data and labelling | The wrong SKU transacted | Duplicate item codes, similar labels, unreadable barcodes |
| Status errors | Unsaleable stock counted as available | Damaged, expired, quarantined or customer-owned goods mixed into the pool |
| Returns timing gaps | Physical stock present before disposition is recorded | Returns area, inspection SLA, restock posting |
| Counting errors | A discrepancy that disappears on recount | Interrupted counts, double counting, non-blind counting, estimated sealed-carton quantities |
The most common failure is correcting the record and closing the count. The record is then right and the process that broke it is untouched, so the variance returns next cycle. Set an escalation ladder in advance: variances within tolerance post with documentation; those beyond a defined threshold trigger an independent blind recount; material variances require supervisor approval, a root-cause code and evidence before any adjustment posts.
Treat every adjustment as a controlled journal entry: original record, physical observation, change amount and unit, a reason code describing a mechanism, operator and approver, and timestamp. Reason codes should read "unconfirmed move" or "wrong unit conversion", not "error" or "miscellaneous", which cannot support prevention. Never rewrite a historical receipt or pick to force current stock into agreement — preserve the failed event and correct it with a traceable transaction.
Stock in a defined process — receiving staging, quality inspection, packed-not-shipped, returns awaiting disposition — is legitimate, but it is also where inventory becomes invisible. Each such location needs a physical area, a named custodian, entry and exit transactions, a maximum dwell time and a reconciliation frequency; without a dwell limit, staging accumulates stock that exists physically and nowhere else, and the next count finds a variance nobody can explain. Returns need particular attention: a unit received before disposition is recorded can legitimately exceed the system balance, so agree a processing window in writing.
Accuracy is both an operating KPI and a financial risk-allocation mechanism, and contracts frequently define it badly. Insist on four things. Define the metric precisely — basis, tolerance, counting population, reporting period and data source — because a percentage without a denominator is not a commitment. Define who counts, how often, and what report you receive, including unresolved investigations rather than just the pass rate.
Then address shrinkage as a commercial term, not a percentage in isolation. Specify the allowance, the measurement base and the reconciliation period together: an allowance measured against throughput rather than inventory value can differ by an order of magnitude on the same operation, and a generous formula is easily neutralised by a short claim notice period or a low liability cap. State whether overages may offset shortages, and press for netting by SKU and facility rather than in aggregate — a surplus of cheap product should never mask a shortage of expensive product. Confirm the valuation basis for recovery, the evidence required to support a claim, and whether you must insure your own goods, since warehouse terms frequently do not cover client inventory by default.
MJDT Global (明杰供应链) treats stock integrity as one chain rather than a warehouse-only concern: clean data at consolidation in Shenzhen, receipt against a pre-advised document set, a count programme agreed at onboarding with defined tiers and tolerances, and variances traced to a reason code before any adjustment posts. Because you cannot walk an overseas floor, we report accuracy on the basis agreed with you, including unresolved investigations, rather than a headline percentage.
Fulfilment sits alongside our regional lanes; see our freight forwarding services and the global network for the full route map.
Accuracy targets, tolerance norms and contract terms differ by market, product category and provider: confirm current definitions with your 3PL and take advice on liability and insurance clauses before signing.
Send your SKU count, monthly order volume, top categories, current warehouse arrangement and any variance reports you already receive. We'll come back within 24 hours with an inventory-control view — a count programme and tolerances appropriate to your catalogue, the root causes we would expect in your category, and the clauses worth fixing in your 3PL agreement.
Get an Inventory Accuracy Review →