When three channels want the same shelf

Somebody is about to lose. In most grocery businesses, the person who decides who loses is a picker with a headset, improvising, forty minutes into a shift, with a promise clock running.

That decision is made thousands of times a day across a large estate, and almost nothing has been written about it. The public conversation about online grocery is still stuck on the building — dark store or supermarket, own fleet or platform. Those arguments are largely settled now. The problem that replaced them is harder, less visible, and considerably more expensive to get wrong.

I want to give it a name, because it does not have one: shared-estate fulfilment. One store network, one inventory pool, one picking workforce, serving several delivery promises at once. And the thing inside it that decides whether the whole model works: promise-window arbitration.

Everyone has quietly converged on the same model

Five years ago the grocery industry ran a very public argument about buildings. It has ended, and the resolution is visible in the reporting of almost every large listed grocer. They are not choosing between fast and scheduled. They are running both, plus everything in between, out of the shops they already had.

Tesco is the clearest case, because it discloses the most. Whoosh, its rapid service, grew 59% in the first half of FY2025/26 and now runs from over 1,600 UK stores, including over 180 large stores.¹ Those large stores are the same buildings that pick scheduled tesco.com orders. Growth did not stop there: Whoosh was up over 30% again in the first quarter of FY2026/27, with total online up 8.9%.² And in July 2026 Tesco began putting Whoosh onto Uber Eats and Deliveroo³ — a fourth promise window, on the same estate, with the same stock.

Walmart has published the number that best captures the shift. In the quarter to October 2025, expedited deliveries — three hours or less — grew 70% and accounted for about 35% of store-fulfilled orders in Walmart US.⁴ Not a separate fast business. A third of the existing store-fulfilled business, now travelling under a much tighter clock.

Ahold Delhaize went further and made the direction structural. It reached e-commerce profitability on a fully allocated basis during 2025, and completed the closure of six US e-commerce fulfilment centres, describing the result as a shift to a "store-first operating model".⁵ By the first quarter of 2026, Ahold Delhaize USA reported record online penetration of 10%, with some banners above 11%.⁶

Then there are the two moves that tell you more than any growth rate. Morrisons launched its own rapid service in October 2025 with a one-hour promise, 158 stores live and 420 targeted by the end of that month — and a cap of 30 items per order.⁷ And Sainsbury's, in February 2026, killed Chop Chop, the standalone rapid app it had run since 2016 across 50 stores, folding rapid delivery into its main app alongside Deliveroo and Uber Eats.⁸

Meanwhile the marketplaces became a channel in their own right rather than an experiment: grocery reached 18% of Deliveroo's group gross transaction value in the first half of 2025, up from 15% a year earlier.⁹ Asda now runs Express Delivery from around 300 superstores through Uber Direct.¹⁰

So the picture is settled. The estate is shared. What almost nobody discusses is what that actually does to the operation.

Four promises, one shelf

Start with the thing a customer never sees: the same physical unit of stock is simultaneously promised to four different clocks, each with its own economics and its own owner of the customer relationship.

The temptation, when you draw this, is to think of it as a portfolio. More windows, more demand captured, same fixed cost. That is the logic that has driven every retailer in the list above, and in the main it is correct — it is the argument I have made at length for store-based quick commerce, and I still believe it.

But it hides an assumption that deserves to be dragged into the light.

Why pooling looks like free money

There is a genuinely beautiful result in operations research, and it is the intellectual foundation of the shared estate whether or not anyone in the building has heard of it.

In 1979, Gary Eppen published a four-page note in Management Science showing that if you consolidate demand across locations, expected holding-and-penalty costs rise only with the square root of the number of demands consolidated, rather than in proportion to them.¹¹ Pool two channels and you need meaningfully less buffer than two separate buffers. Pool four and the saving is larger again. Variability partially cancels: when rapid demand spikes, scheduled demand is often quiet, and one pile of stock covers both.

This is why serving four promise windows from one shelf feels like something for nothing. You are not building four inventories. You are building one, and it absorbs the wobble of all four.

Eppen's result is real. But it comes with conditions, and the conditions are precisely where grocery delivery lives.

And why the maths quietly stops working

Eppen's proof assumes demands that are identical and uncorrelated, and approximately normally distributed. Hold that against what actually happens in a supermarket at six in the evening.

Rapid-delivery demand is not normally distributed. It is spiky, it concentrates on a narrow set of SKUs, and it is positively correlated across channels — the rapid customer, the scheduled picker and the walk-in shopper all want milk, bread, chicken and bananas, and they all want them in the same three-hour window each evening. Correlated demand is exactly the case where pooling gives back its advantage.

The formal version of the problem is worse than intuition suggests. Bimpikis and Markakis showed in 2016 that when demand is heavy-tailed rather than well-behaved, the benefit of pooling scales as n^(α−1)/α, where α is the tail index.¹² At α = 2 you recover Eppen's square root. As α falls towards 1 — heavier tails, spikier demand — the benefit flattens towards nothing.

Bimpikis and Markakis give one worked illustration of how badly this can mislead: at α = 1.1, a manager who wrongly assumed exponentially distributed demand would overestimate the absolute benefit of pooling by a factor of 2.7.¹² That is a single example rather than a general law — but the direction of the error is the point. Pooling models flatter the shared estate. The flattery is largest exactly where demand is spikiest, which is exactly where the twenty-minute promise lives.

None of this argues against the shared estate. It argues against assuming the shared estate is self-organising. The buffer that the theory says you can remove is the buffer that was quietly absorbing your arbitration failures.

The phantom unit

Now add the problem that makes the whole thing concrete, and that anyone who has run a store already knows in their bones: the system's idea of what is on the shelf is frequently wrong.

The definitive study here is DeHoratius and Raman, who examined nearly 370,000 inventory records across 37 retail stores and found 65% of them inaccurate.¹³ Not 6.5%. Sixty-five. Recent grocery-specific work reports the same order of magnitude, with roughly 39% of SKUs showing negative record inaccuracy — the system believing there is stock when the shelf is empty — and store-wide sales lifting around 11% after audits, with the entire benefit concentrated on those negative-inaccuracy items.¹⁴ Perishables, the categories that matter most to a grocery delivery basket, benefit around three times more from auditing than ambient lines.

This is the mechanism that turns a data problem into a customer problem. A phantom unit sits in the system. It gets promised. And of the four channels queuing for it, the one that discovers it does not exist is almost always the fastest one, because the twenty-minute order is the one being picked right now.

What the customer actually experiences

The customer does not see inventory record inaccuracy. They see a different bag than the one they ordered.

Which? surveyed 3,552 UK adults about 1,417 online grocery shops, with fieldwork in October and November 2025. One in five shoppers received a substitution in their most recent order — 20%, improved from 29% the year before.¹⁵ The spread across retailers is wide enough to be interesting on its own.

And substitution is not a neutral event. Research published in the Journal of Retailing found that retailers currently achieve only 55–70% acceptance of the substitutes they offer, and that acceptance moves sharply with how the substitute is chosen: matching flavour lifts acceptance by 26–30% in horizontally differentiated categories, matching brand by 18–19% in vertically differentiated ones, and offering something the customer has bought before by 18–32%.¹⁶

Sit with that range. Roughly a third of substitutions are rejected outright — and the difference between a well-chosen substitute and a lazy one is measured in double-digit percentage points of acceptance. This is one of the few places in grocery operations where a piece of logic, well built, is worth more than a piece of capital.

For context on the underlying difficulty: the classic worldwide study of physical shelf availability, covering 52 studies and 71,000 consumers, put the average out-of-stock rate at 8.3%.¹⁷ That is the raw material every online promise is built on top of.

The congestion tax nobody budgets for

There is a second cost of the shared estate, and it is spatial rather than informational. Pickers and shoppers occupy the same aisles at the same time, and the peak of online demand sits inside the peak of store footfall.

This has now been studied properly. Neves-Moreira and Amorim modelled the dynamic in-store picker routing problem with a major European retailer across six store configurations, and found that learned picking policies reduced customer encounters by up to 50% compared with policies that optimised only for picking.¹⁸ The work is simulation rather than field data, and worth citing carefully for that reason — but the finding is directionally important, and it cuts against the usual instinct. Optimising purely for pick speed makes the store worse for the people standing in it. There is a real trade, and it can be managed deliberately rather than discovered through complaints.

Who loses the last unit

Which brings us back to the chicken thighs.

Every retailer running a shared estate has an answer to the arbitration question, whether or not they have written it down. Usually they have not written it down, which means the answer is being improvised at shelf level by whoever gets there first. That is not a rule. That is the absence of one.

The academic work here is thin but pointed. Hofstra and Spiliotopoulou ran incentivised experiments on how people ration scarce stock between a safe channel and a riskier, higher-margin one. Their finding: when profit differences between channels are medium to large, people allocate significantly less inventory to the risky, more profitable channel than expected-profit maximisation would prescribe.²⁰ It is a lab study, not field practice, and the effect is conditional rather than universal — but it describes something recognisable. Left to human judgement under time pressure, allocation drifts towards the familiar and the safe, not towards the profitable.

The practical consequence is that arbitration happens either by design or by default, and by default it optimises for whoever is shouting loudest — usually the channel with the shortest clock, which is frequently the channel with the thinnest basket.

The row that matters is the last one. The first four are legitimate strategic choices; different businesses will pick differently and be right. The fifth is what happens when nobody has chosen, and it is the most common state of affairs in the industry.

The channel that changes the rules

Marketplace orders deserve their own treatment, because they break an assumption the other three share: that you own the customer.

On a marketplace order, someone else owns the interface, the notification, the complaint and the reorder. You own the shelf, the pick and the blame. Deliveroo's grocery share reaching 18% of group GTV⁹ is a large number for a channel where the retailer is, in customer terms, a supplier.

The published economics deserve a caveat that most commentary skips. Uber Eats and DoorDash both publish marketplace commission tiers — roughly 15% to 30% of order value depending on the plan.²¹²² Those are restaurant-facing rate cards, and both companies state plainly that large-merchant rates are negotiated individually. I could not find a single credible source tracing a named large grocer to a specific commission rate, and anyone quoting one at you should be asked for the source.

What is disclosed is suggestive in a different way. DoorDash's chief financial officer told the Q4 2025 call: "I expect our entire retail and grocery business to be unit economic positive in the second half of the year."²³ The inference — that grocery and retail were not yet unit-economic positive for the platform at that point — is mine rather than the company's, and should be read as inference. But it sets up the question every grocer should be asking: if the platform is still working towards positive unit economics on grocery, and the platform takes a share of every order, what exactly is the retailer's contribution margin on that same order, and is anyone measuring it separately from own-app orders?

The honest answer, industry-wide, is that nobody publishes it. Which is itself informative.

Six rules that survive contact with a real estate

None of the above argues for fewer channels. Demand is demand, and a store estate that can serve four windows is more valuable than one that can serve two. But shared estates do not self-organise, and the difference between a good one and a bad one comes down to whether these six things have been decided deliberately.

1. Write the arbitration rule down, before you need it. Which channel loses the last unit is a strategic choice with a defensible answer. Decide it, publish it internally, and make it a system behaviour rather than a picker's judgement call at 6pm. Any rule beats no rule, because a rule can be measured and improved and improvisation cannot.
2. Do not give every channel the same catalogue. The fastest promise should carry the tightest range. Every additional line in a twenty-minute assortment lengthens the pick path for every order in that window, including the ones that do not contain it — and widens the surface on which inventory inaccuracy can bite. Morrisons' 30-item cap is a blunt version of the same instinct.
3. Buffer the fast channel deliberately, not accidentally. Pooling theory says you can hold less. The tail behaviour of rapid demand says the last few units of the top hundred SKUs are worth protecting anyway. Treat a small, explicit, SKU-limited reserve as the cost of keeping the shortest promise credible, and measure what it costs you.
4. Make substitution logic a first-class product, not a fallback. With acceptance at 55–70% and the levers well documented — flavour, brand, previously purchased — this is one of the highest-return pieces of logic in the operation, and one of the cheapest to improve.
5. Attack negative record inaccuracy where it hurts most. Audit perishables and top-velocity lines far more often than the rest. That is where phantom units are created and where the audit return is highest. Every phantom unit removed is a substitution that never happens on the tightest clock.
6. Measure the tail, per channel, never the average. Average delivery time and blended availability hide precisely the failures that matter. The metric that predicts whether a shared estate is working is the proportion of orders that miss their promise, broken out by channel, with the reason attached. Blended numbers improve while the fast channel quietly rots.

What is not public

Several things this argument would benefit from simply are not disclosed by anyone, and it is worth naming them rather than papering over them.

No retailer publishes its channel-priority rules. Not one. Available-to-promise tie-break logic is treated as competitively sensitive, so there is no public benchmark for how the last unit should be allocated. Everything above is reasoning from mechanism and from the few adjacent studies that exist.

No retailer has publicly admitted, in a quotable line, that serving several promise windows from one store creates operational conflict. The evidence is circumstantial rather than confessional: Morrisons' item cap, Sainsbury's consolidation of Chop Chop into one app, Murphy's remark about capacity filling as fast as it is added. Circumstantial evidence from four large retailers pointing the same way is worth something, but it is not an admission and I will not present it as one.

Nobody discloses the margin gap between own-app and marketplace orders. Albertsons has said its digital operations carry "a considerably lower gross margin rate" than traditional grocery,²⁴ but that is digital versus store, not own-app versus platform.

In-store picking productivity versus dedicated fulfilment picking is vendor data only. I found no peer-reviewed or primary-filing comparison of units per hour. Anyone quoting a clean ratio is quoting marketing material.

And there is no citable figure linking substitution rates to repeat purchase, basket value or NPS. This is the one that surprised me most. The claim is asserted constantly. Kroger's own analytics arm markets an analysis of exactly this question and publishes no numbers. If you have seen a real source, I would genuinely like the reference — and if you have not, be careful of the number you are about to repeat.

Common questions

Is shared-estate fulfilment just omnichannel with a new name?

No. Omnichannel describes what the customer can do. Shared-estate fulfilment describes what the operation must resolve: several promise clocks drawing on one pool of stock, one workforce and one dispatch point. The customer-facing word has existed for a decade; the operating problem underneath it has been left largely undescribed.

Should a fast channel have ring-fenced stock?

A small, explicit, SKU-limited reserve on top-velocity lines usually earns its keep, because the fast promise is the one that breaks most visibly. Wholesale ring-fencing does not — it recreates the separate inventories the shared estate exists to avoid, and reintroduces exactly the waste that killed the dark-store generation.

Does adding a marketplace channel cannibalise your own app?

Unknown, publicly. Platforms claim high incrementality; none has published a methodology, a denominator or an independent verification, and no retailer-side study with numbers exists in the public domain. Treat every incrementality claim you are shown as a hypothesis to test on your own data, not a finding.

What is the first thing to fix on an estate already running three or four windows?

Reporting, before operations. Split promise-miss rate by channel with reasons attached. Most estates cannot currently answer "which channel is absorbing our availability failures", and until that is visible, every fix is guesswork.

Does this get easier or harder at scale?

Harder, and in a specific way. At ten stores a shared estate runs on the goodwill and judgement of a few strong managers. At sixty, judgement has to become a system, because goodwill does not replicate — and arbitration is the part most often left as judgement longest.

Sources