Manual Audits Are Already Too Late. That Gap Is Your Margin.

There is a particular kind of quiet in a store after close on count night.

The doors are locked. Somebody is working an aisle with a scanner and a clipboard, calling numbers to somebody else who is writing them down. It takes hours. It costs payroll. And at the end of it you get a variance report that tells you, with real precision, that the system was wrong.

Then everybody goes home, the records get corrected, and the next morning the store opens and the whole thing starts again.

Most store owners read that variance report as the finding. It is not the finding. It is the receipt.

The count is not the problem. The count is the aftermath.

By the time a physical count catches a discrepancy, the discrepancy has been sitting there for weeks. Reorders went out in that window against numbers nobody had checked. Availability got promised to customers in that window against numbers nobody had checked. Hours got spent looking for product that was not where the record said it was.

The count does not create that loss. The count reveals it, late, after it has finished happening.

This is Inventory Truth Decay (ITD), the continuous process by which a record drifts away from what is physically in the building. ITD is not an event. It does not happen to you on a Tuesday. It runs the entire time the store is open, because the store keeps changing and the record does not change with it.

Manual audits are not the villain here. I want to be exact about that, because the conclusion people jump to is the wrong one. Audits are not what is hurting you. Audits are a late response to something that was already hurting you. Counting less would make it worse. Counting more does not make it stop.

Someone measured how fast it decays. The design is the whole story.

In 2019, researchers Yacine Rekik, Aris Syntetos and Christoph Glock published a study through the ECR Community that ran an experiment at one retailer across roughly 27,000 SKUs.

The design is what makes it worth your attention.

They began with a stocktake and removed the inaccuracies they found from the system, so the records started from a corrected position. In all they ran twelve stocktakes, counting that first one: six, a four-week break, then six more. And through every count after the first, in the report's own words, the inaccuracies they found "were recorded, but not removed from the system."

So the discrepancies those eleven counts turned up were deliberately left in place. What the researchers withheld was stocktake correction. The store kept trading the whole time, and the report does not claim that every other kind of inventory adjustment stopped.

The report charts the result across stock counts 2 through 12. Over that span the share of SKUs with accurate records falls from 39.89 percent to 20.75 percent.

That is not a story about counting failing. Those eleven counts were not fixing anything, by design. It is something more useful: a measurement of how far inventory records drift once periodic reconciliation is taken out of the loop, observed in a store that stayed open throughout.

It is the number the retailer in that same research said nobody had:

"We don't know how inventory records deteriorate in our company. Does it reach 74% a week after the stocktake, or does it degrade over the next months?"

That is a retailer admitting they do not know the shape of their own loss curve. They know the records go bad. They do not know how fast. Nothing in their stack is watching between the counts.

The errors are not symmetrical

The same study found something about the direction of the drift that matters operationally.

In the report's words: "negative inaccuracies clearly accumulate over time; positive discrepancies, however, seem to remain relatively stable." The researchers attribute this to positive errors often being temporary misplacements, while negative errors are frequently permanent loss.

Both cost money, and they cost it differently. The report is specific. When the system claims more than exists, "orders are placed too late, which increases the risk of a stockout." When the system claims less than exists, "orders are placed too early, leading to unnecessary inventory in the system that causes extra costs."

In this experiment, the error that empties the aisle compounded with time and the error that fills your stockroom with product you already had stayed relatively flat. Both still cost money. The researchers were careful not to turn that into a law. They wrote that the result "clearly motivates the need for further scientific research on the drivers and behaviour of negative and positive discrepancies."

What the gap is worth

The 2019 research also found that correcting inventory records lifted sales at the participating retailers by 4 to 8 percent, and by more than 14 percent on what the report calls high discrepancy products.

Those are results from the stores in that study, not a promise of what any particular store would see. But the direction of the finding is the point. Not stocking more. Not discounting. Making the record match what is actually in the building.

That number is pricing the gap. It says a meaningful share of what looks like a demand problem, or a merchandising problem, or a bad location, may be a record problem wearing a disguise.

And most of this industry is small. According to NACS, 95,672 US convenience stores, about 63 percent of the industry, are owned by companies running ten stores or fewer. Those counts do not on their own prove that these smaller companies carry thinner margins or worse record accuracy than national chains, and I am not going to claim they do. What is true is that the same decay runs in a store where counting is one person, a scanner and a night, with no inventory team to absorb it.

Why counting harder does not close it

The instinct, once you see the slope, is to count more often. Halve the interval, halve the gap.

It helps. It does not close it, for three reasons.

First, more counting costs more. Exactly how much more depends on your scope, your staffing and your method, so I will not pretend there is a clean multiplier. But the cost of counting rises with frequency, and the accuracy you buy does not rise with it in a straight line, because you are sampling a continuous process at discrete points and the process resumes the moment you stop.

Second, counting introduces its own errors. A count is performed by people, at night, at speed, against a list. Some portion of every variance report is the count being wrong rather than the record being wrong, and nothing in the process tells you which is which.

Third, and this is the one that ends the argument: you cannot shrink the interval to zero. However often you count, there is a window between counts where nothing is watching. You can make the window smaller. You cannot make it not exist. And the loss lives in the window, not in the count.

The honest version of the newer answer

There is a serious attempt to solve this with modeling, and it deserves describing accurately rather than dismissing.

ECR Retail Loss published work in 2026 drawing on roughly 1.3 million stock-audit observations across six grocery retailers. At one of the retailers studied, the incidence of records showing stock that is not physically there ran at about 18 percent where counting happened monthly, against more than 27 percent where auditing happened twice a year. That is one retailer, not a universal law, and the research does not claim the same relationship holds in every store. It is still a real signal that the interval matters.

The same body of work can rank which items are most likely to have gone wrong, apply low-risk corrections automatically, flag borderline cases for a quick check and send staff to count the rest. That is a genuine advance. The economic gains attached to it in that report are model-based estimates rather than results measured in the field, and the report says so itself. Treat the direction as real and the size of the prize as unproven.

That approach is, structurally, a model of the record. It learns the shape of past errors from audit history and transaction data, then projects that shape forward. It is good at telling you which records have probably gone wrong. Nothing in it has looked at the product since the last time a person did. I am describing this method, not every predictive system on the market.

A prediction about a record is still a statement about the record. It is not an observation of what is there.

Where this leaves the store owner

The question most inventory conversations start with is "what does the system say."

The more useful question, and the one almost nobody has a way to answer, is: how much should I trust what the system says, right now, today, given how long it has been since anyone actually looked?

Every record in your store has an age. Some are fresh. Some were accurate three weeks ago and have been quietly wrong ever since. Some are technically correct and operationally useless, because the item is in the building and nobody can find it. Right now, for most store owners, all of those records look identical on the screen.

That is the gap. Not the variance on count night. The weeks before it, where every number looks equally confident and some of them are already lying.

The count is the receipt. That gap is your margin, and it has been running the whole time.

What would it change about how you order, staff and promise availability if every number on your screen came with a confidence level attached?

---

Sources

1. Rekik, Y., Syntetos, A. and Glock, C. Inventory Inaccuracy in Retailing: Does it Matter? ECR Community Shrinkage and On-shelf Availability Group, October 2019. cardiff.ac.uk
2. ECR Retail Loss. Predicting Inventory Record Inaccuracy. 2026. ecrloss.com
3. NACS. Convenience Store Industry Store Count. convenience.org