Amazon Inventory Management

Inventory is what lets you grow and what quietly eats the margin you grew for. The whole discipline is two numbers: the share of your traffic that sees a buyable offer with the delivery promise intact, and what holding inventory costs against what it sells. This is how an in-stock manager keeps both healthy, and what to do when either one breaks.

Amazon charges you for holding too much. Since April 2024 it also charges you for holding too little. The platform itself treats inventory as a discipline with two failure modes, and it priced both.

The heavy side is familiar: monthly storage billed on the cubic feet you occupy, a utilization surcharge that ladders up as your stored volume outgrows your shipped volume, and an aged-inventory surcharge once units pass 181 days in the network.

The lean side is newer and less read. The low-inventory-level fee applies to standard-size products that carry consistently low inventory relative to unit sales. Amazon's stated reason: thin inventory stops it spreading your stock across the network, which degrades delivery speeds and raises its shipping costs. Its stated avoidance: hold more than four weeks of inventory relative to sales.

Read the two columns as one instruction. Amazon is telling you, through its fee schedule, that it wants you between roughly four weeks of cover and roughly twenty-two, and it charges an escalating toll on either side of that band.

Carrying too much costs money slowly and recoverably. Running out costs rank, and rank does not come back on its own.

This asymmetry is the single most useful idea in inventory planning, and almost every published rule of thumb ignores it.

Overstock is a bad outcome with a floor. You are out the storage, the surcharges and some margin, and you can usually see the whole bill. The units still exist and most of the money is recoverable.

A stockout has no such floor. You lose the orders you could not fill, which is the part people count. Then the listing stops converting, the flywheel unwinds, and the position gets taken by whoever is still available. Coming back means buying velocity again at whatever the shelf now charges for it.

The goals of this discipline fit on one dashboard: instock% and storage as a share of net sales. Everything else in this guide exists to move one of those two numbers before it moves on its own.

One more wrinkle sits under every rung: availability has a middle state. A listing can be technically buyable, with every unit stuck in transfer between fulfilment centres, quoting a delivery date that loses the shopper. Available on paper. Losing the sale in practice.

Read together, the two KPIs describe the whole job. One measures whether the gate is open when demand shows up. The other measures what keeping it open costs.

A seasonal demand curve is built from three signals: your own sales history, Amazon search volume, and general search interest. The skill is not blending them. It is knowing when each one lies.

Your own sales are the truest signal about your own products, and the most contaminated one: every stockout, promotion and ad step-up in your history is baked into the shape. Amazon search volume reads category intent on the platform where the buying happens, but the usable history is short. General search interest runs five years deep and is free, but it measures looking, not buying.

The weights are a judgment about data quality, per product, and the honest way to show what that judgment looks like is real decisions. These five are from the same operator's planning workbooks, built in the same month, for five different products.

Same operator, same month, and the weights run from one hundred percent own-sales to zero. That range is the lesson. A fixed blending formula would have been wrong four times out of five.

A cover target is how many weeks of forward demand you intend to hold. One target across a catalogue is the standard mistake: the right answer is a spread, set by pricing a stockout per product.

The spread exists because the cost of being wrong is not the same on every product. On the SKU carrying your rank, your reviews and your brand search, a stockout is expensive in a way that does not show up for months. On a long-tail variant, a stockout is a lost order and very little else.

So the products whose position you cannot afford to lose carry more, and the tail carries less. As an example of the shape: sixteen weeks on the products that carry your rank, twelve on the core, eight on the tail. An example, not a prescription. The spread is a risk-management call per product, priced by the arithmetic below.

The arithmetic is a comparison almost nobody runs: what one stockout event actually costs, against what a year of carrying the extra cover costs. You cannot set a defensible cover target without pricing a stockout, so price one.

Run it once for a hero product and once for a tail variant, and the spread stops being advice and becomes your own numbers. The hero justifies weeks of extra cover with room to spare. The tail usually does not, and that is the spread.

Order up to a level, on a fixed cadence, with math a director can check by hand. There is no model in this chapter, deliberately.

First the cadence. If you reorder the moment cover drops below target, you will be back below target a week later, and you will place a small order every week forever. Suppliers price against that, and your team stops believing the plan.

The fix is periodic review. Decide how often you actually place orders, then order enough to carry you to the next order rather than enough to touch the target. The buy is bigger, there are fewer of them, and each one is defensible to a supplier, a finance team and a warehouse.

Then the arithmetic. All of it fits in four lines.

Three rules keep the arithmetic honest, and all three come from the same discipline: the plan must describe reality, not repair it.

A forecast you never score is an opinion. Score it monthly, and score it for bias, because bias is the error you can fix.

On a seasonal product the storage bill arrives before the revenue does. That spike is not a mistake. It is the plan, and the discipline is budgeting it rather than panicking at it.

Watch storage as a share of net sales and a flat-demand product is easy: the ratio runs level all year, and any movement is a signal.

A fourth-quarter product is a different animal. Inventory has to land in August and September, ahead of an October peak, so for two months you pay storage on a warehouse full of product while sales run at their annual low. The ratio spikes. Nothing is wrong.

The failure mode is not the spike. It is a team that never planned the spike, sees the ratio triple in September, and reacts: a panicked discount into the pre-season lull, or a removal order on stock the season would have sold. The overcorrection costs more than the bulge ever would.

The removal fee is the small half of the cost of a removal. Amazon does not send back a pallet.

Removal reads as the responsible option. You keep the units, you pay a modest per-unit fee, and you resell them somewhere else. Every published guide quotes that fee and stops there.

What actually arrives is not a pallet. It is units returning piecemeal, in many mixed boxes, over weeks, from more than one location, in no particular order, with no manifest that matches how you would want to receive them.

Somebody at your warehouse then opens all of it, sorts it, checks condition and puts it away. For most brands that labour is larger than the removal fee, sometimes by a multiple, and it lands as disruption to a team that had other work.

A price cut costs you the margin, and it also teaches the shelf a price. The second cost outlasts the promotion.

Discounting to clear excess is the most common route and the one whose full cost is least examined. The margin arithmetic is easy and immediate. The rest is not.

In our own measurement of competitor price moves, cuts that were never restored are associated with a worse fourteen-day share outcome than temporary promotions that went back up. A permanent drop does not read to the market as a deal. It reads as a new price, or as distress.

A higher price rations the remaining units towards the buyers willing to pay for them, and recovers margin on stock you cannot replace yet.

It does three things at once. It slows the sell-through, which buys time for the replenishment to land. It improves the margin on every remaining unit. And it does both without turning the listing off, which matters for the reasons in the next three chapters.

The limit is that a large increase on a price-sensitive shelf can cost conversion badly enough to hurt rank on its own, which is the thing you were protecting. Move it far enough to ration, not far enough to look like a different product.

Cut defensive and branded spend first. Protect acquisition, but only while you will still have stock when those customers arrive.

The instinct when stock is short is to pause everything. That is cheaper than doing nothing and worse than triaging.

These rules sit beside the wider argument about what advertising is actually buying, which is covered in the advertising guide.

Recovery is not the fall in reverse. You buy the position back at whatever the shelf charges now.

Restocking does not restore rank. It restores availability, which is the precondition for earning rank again, at a moment when a competitor has had weeks of uncontested velocity.

What follows is a rebuild: spend to regain the placements, at a conversion rate that has to be re-established, against a rival whose review count grew while yours did not. It takes sustained weeks of selling to re-teach the algorithm what your normal velocity is. That spend is real, it lands after the stockout has stopped being visible in the numbers, and it is almost never attributed back to the shortage that caused it.