Grocery Route Optimization: Why You Are Planning Routes for an Order That Has Not Been Picked Yet

Grocery route optimization sequences orders into vehicle plans using the weight, volume and temperature class of what the customer ordered. In pick-from-store and micro-fulfillment models, picking happens after that plan is built. So the manifest the optimizer solved and the manifest the vehicle loads are two different things, separated by every short-pick and substitution the picker made in between.

This is not a data-quality problem to be tidied up. It is the normal operating condition of the category, and it is a structural difference from parcel or freight, where what was ordered is what ships. The question worth answering is not how to eliminate the gap. It is which dimensions of the plan can absorb it and which ones break, because those two get treated identically today and only one of them matters.

Why the gap exists and how large it is

The stockout rate that drives substitution is well established. McKinsey’s work on omnichannel grocery fulfillment, citing IRI’s CPG Supply Index, notes that 11% of edible packaged consumer products are out of stock in store, and that in a pick-from-store model this translates to 10% to 15% of orders experiencing a stockout and potentially requiring substitution during fulfillment.

The causes sit upstream of delivery, which is worth stating because it sets what a routing platform can and cannot fix. Research by Corsten and Gruen synthesizing more than 50 studies, summarized by ECR Retail Loss, puts the global out-of-stock rate at 8.3% and attributes causes across five work processes: store stocking 38%, store forecasting 22%, planning 11%, store ordering 11% and supply 9%. Delivery execution is the 9%. Routing does not reduce substitution. It has to survive it.

Substitution is also not a neutral event for the customer. Peer-reviewed work on substitution policy in online grocery retailing examines how substitution handling shapes customer response, which matters operationally because approval and refusal both consume time at or before the door.

The consequence of getting the discontinuous case wrong is also asymmetric in a way it is not for durable goods. ReFED’s U.S. Food Waste Report puts 2024 surplus food at 70 million tons, roughly 29% of the US food supply. When a temperature-class break spoils product, the cost is the goods plus the redelivery plus the customer contact, where a failed parcel delivery is a redelivery alone. That asymmetry is the reason the binary dimensions deserve disproportionate design attention.

And the margin structure means this cannot be solved by spending. FMI’s Food Retailing Industry Speaks 2025 puts food retail net profit margins at 1.7%. A category running on 1.7 points cannot absorb a structural inefficiency by adding vehicles, which is why the response has to be a better plan rather than more capacity.

How to quantify and design for the plan-to-pick gap

1. Compute the probability at route level, not order level

An order-level substitution rate sounds manageable. Aggregated across a route it stops being an exception case.

At 40 orders per route and a mid-range 12.5% rate, the expected number of changed orders is five and the probability that the route loads exactly as planned is 0.5%. Every route is a changed route. Any process that treats manifest mutation as an exception workflow is treating the normal case as an exception.

2. Check the weight effect, then stop worrying about it

This is where most operators expect the damage, and the arithmetic says otherwise. Short-picks only subtract, so a route arrives systematically lighter than planned rather than heavier. If 12.5% of orders lose one line worth roughly a twenty-fifth of order weight, route weight falls about 0.5%, and a route planned to 90% of capacity arrives at 89.5%.

That is a genuinely useful negative result. It means the plan-to-pick gap is not a capacity risk, so building weight headroom against it is wasted capacity, and re-planning a route because the manifest changed by half a percent is churn without benefit.

3. Find the dimensions that fail discontinuously

Weight and volume degrade gracefully. Temperature class does not. A chilled item substituted for an ambient one, or the reverse, changes which compartment the order needs and where in the sequence it can sit. If a fifth of substitutions cross a temperature boundary, on a 40-order route with five changed orders the expected number of boundary crossings is one, and the probability of at least one is 63%.

A route planned as ambient-only that acquires one chilled line is not marginally worse. It is either infeasible or it degrades the product, and no amount of weight headroom addresses it.

4. Understand how basket size multiplies a small line-level rate

Order-level rates hide the mechanism. If each line has an independent probability of change, the chance a basket arrives fully intact falls exponentially with the number of lines.

This is an illustrative model rather than a published result, and real substitutions are correlated rather than independent, so treat it as a sensitivity rather than a forecast. The direction is what matters: large-basket weekly-shop operations face a materially different exposure from small-basket convenience operations at the same line-level rate, and a single routing configuration serves them differently.

5. Price the service-time change, not just the load change

A substitution that needs customer approval consumes time somewhere: a pre-delivery contact, a conversation at the door, or a refusal that becomes an unplanned return on the vehicle. That is a change to the service-time estimate the route was built on, and service time is the larger share of a grocery route’s clock. A composition change that adds two minutes at five stops costs more route time than the entire weight effect costs capacity.

Refusal is the tail case and it has a second-order effect. A rejected substitution does not simply end the transaction: the item travels back on the vehicle, so a route that departed with no planned return volume acquires some, and if returnable crates are involved the reverse capacity was never modeled at all. That is a small volume on any one route and a systematic one across a fleet, and it is the one dimension where the plan-to-pick gap and the reverse-logistics gap compound rather than sitting in separate reports.

6. Decide re-plan against tolerance deliberately

Two responses exist and they are not interchangeable. Re-planning after picking completes produces a feasible plan at the cost of churn, driver disruption and a later departure. Building tolerance into the original plan, by reserving compartment flexibility and service-time buffer where composition is most volatile, costs a little capacity on every route and no disruption on any of them. Given that essentially every route changes and the weight effect is negligible, tolerance is the better default and re-planning should be reserved for the discontinuous breaks.

7. Sequence picking and loading against the route, not the aisle

If loading order does not match reverse stop order, the driver re-sequences on the dock or at the doorstep. Substitution makes this worse, because a late-picked replacement lands wherever there is room rather than where the sequence needs it. The plan-to-pick gap therefore shows up as handling time even when the manifest is otherwise fine.

Where grocery differs from adjacent categories

The middle column is the one most grocery operators actually run and the one least well served by a routing configuration designed for the first or third.

Five criteria for evaluating grocery routing against composition change

1. Can the plan carry compartment tolerance as an explicit input? Ask whether you can reserve chilled or frozen capacity on a route that has no chilled orders at plan time, and what that reservation costs in the objective function.

2. Does re-optimization trigger on composition, not just on time and location? A manifest change that breaks a temperature constraint should trigger a re-plan. A half-percent weight change should not. Ask what the trigger conditions are and whether they are configurable.

3. Is service time modeled per order rather than per stop type? Orders with substitutions take longer. If service time is a flat allowance, the route will run late for a reason the plan cannot see.

4. Does loading sequence derive from the route? Confirm the pick and load instruction is generated in reverse stop order and that a late substitution is placed accordingly rather than appended.

5. Can it separate the substitution decision from the routing decision? Which replacement to offer is a merchandising and customer decision. Whether the resulting order still fits the plan is a routing decision. A platform that conflates them will make the wrong one badly. The test is whether a substitution can be approved commercially and then separately rejected by the plan as infeasible, with the conflict surfaced rather than silently resolved in favor of whichever system wrote last.

What this looks like in enterprise deployments

A grocery brand delivering fresh and perishable goods to homes across more than 30 cities through contracted third parties shows the pattern under real conditions. Before, capacity was booked per carrier with no unified view of cost or service, so a manifest change had nowhere to go except a phone call. Locus orchestrated carrier allocation across those partners, producing 33% faster deliveries, 15% lower fulfillment cost, 25% less manual shipping time and customer support resolution 10 to 20 times faster. The support figure is the one that speaks to composition change, because a substituted or short-picked order is disproportionately likely to generate the contact that resolution time measures.

A leading North American retailer running multi-hundred stores across ocean, rail and road replaced six legacy systems with a single orchestration layer, reaching over 99% on-time store delivery with route compliance above 95% and exceptions resolved in under two hours. Route compliance is the relevant metric here: when the loaded manifest differs from the planned one, compliance measures whether the plan was still worth following, and a plan built with tolerance is one drivers can keep.

Four mistakes in grocery route planning

Treating manifest change as an exception workflow. At normal substitution rates it happens on essentially every route. An exception process that fires 99.5% of the time is the main process.

Building weight headroom against substitution. The weight effect is roughly half a percent and directionally light. Headroom reserved for it is capacity given away for nothing.

Re-planning on any composition change. Churn has a real cost in driver disruption and departure delay. Re-plan on the discontinuous breaks, tolerate the rest.

Using one routing configuration for large-basket and small-basket operations. A 60-line weekly shop and a 10-line convenience order carry very different intact probabilities at the same line-level rate, and they need different tolerance settings.

How Locus plans for orders that change after planning

Locus, the world’s first Decision-Intelligent, Agentic TMS, addresses this through constraint fidelity and continuous re-decisioning rather than through a separate exception process. The route planning system sequences against more than 250 real-world operating constraints, and the ones that carry this problem are vehicle capacity by weight, volume and temperature zone together, fleet mix across refrigerated, ambient and two-wheeler capacity, cold chain sequencing, and stop-level time windows. Holding temperature zone as a first-class capacity dimension alongside weight and volume is what allows a compartment break to be detected as a constraint violation rather than discovered at the vehicle.

Because plans are produced in roughly two minutes and re-optimized continuously as execution events arrive, a manifest change that does break feasibility can be absorbed by re-sequencing rather than by holding the route. That speed is what makes selective re-planning practical: the discontinuous breaks can trigger a re-plan without the whole shift waiting for a batch cycle.

The agentic layer decides which changes warrant a person. Autonomy Levels run per agent and per domain, so a composition change inside tolerance can be absorbed autonomously while a temperature-class break routes to a dispatcher, and Explainability and Traceability record the trigger, context, reasoning, action and outcome for each of those decisions. That is the mechanism that keeps escalation volume proportionate to the changes that actually matter rather than to the number of changes that occur.

Across grocery and e-grocery deployments, route planning delivers up to 34% fewer miles, 25% higher drop density and 28% fewer trips through order consolidation, with fleet utilization up to 90% and 99% on-time delivery.

Locus supports more than 360 enterprise customers across 30-plus countries, with over 1.5 billion deliveries optimized, more than $320 million in documented client logistics savings and 99.99% uptime. It has been recognized by Gartner for seven consecutive years, featured in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies, named a Leader in TMS by QKS Group (SPARK Matrix), and ranked #1 in Route Planning on G2’s 2026 Best Software Awards.

In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

So how should grocery route optimization handle orders that change after the plan is built? By accepting that essentially every route changes and designing for the dimensions that break discontinuously rather than the ones that drift. At a 12.5% order-level substitution rate and 40 orders per route, 99.5% of routes carry a changed order, yet the weight effect is only around half a percent, so capacity is not the exposure. Temperature class is: with a fifth of substitutions crossing a boundary, 63% of routes acquire an unplanned temperature class, and that is a feasibility break rather than a degradation. Locus holds temperature zone as a capacity dimension alongside weight and volume across more than 250 operating constraints, re-optimizes in roughly two minutes so selective re-planning is practical, and uses per-domain autonomy levels so a tolerable change is absorbed while a compartment break reaches a dispatcher. For grocery operators running pick-from-store or micro-fulfillment at scale, planning for the manifest that will load rather than the one that was ordered is the difference between a plan drivers follow and a plan they rebuild. Request a Locus route planning assessment to model your own plan-to-pick gap.

Frequently Asked Questions

How often do grocery orders change between routing and loading? More often than route-level reporting suggests. McKinsey’s omnichannel grocery work, citing IRI’s CPG Supply Index, notes 11% of edible packaged products are out of stock in store and 10% to 15% of pick-from-store orders experience a stockout. Aggregated across a 40-order route at a 12.5% rate, the probability that at least one order has changed is 99.5%.

Does substitution create a capacity problem for the route? Rarely. Short-picks only subtract, so routes arrive systematically lighter, and the effect is around half a percent of planned weight. A route planned to 90% of capacity arrives near 89.5%. Reserving weight headroom against substitution gives away capacity for a risk that is not material.

What actually breaks when an order changes after planning? Temperature class and service time. A substitution that crosses a temperature boundary changes which compartment the order needs and where it can sit in the sequence, and at a 20% crossing rate roughly 63% of routes acquire an unplanned temperature class. Substitutions that need customer approval also lengthen service time, which is the larger share of a grocery route’s clock.

Should the route be re-planned after picking completes? Selectively. Re-planning every route because the manifest changed produces churn on a near-universal event, and the weight drift does not justify it. Re-plan when a change breaks a hard constraint such as temperature class or window feasibility, and build tolerance into the original plan for everything else.

Does basket size change the exposure? Substantially. If line-level change is roughly independent, a 40-line basket at a 2% line rate arrives fully intact about 45% of the time against 82% for a 10-line basket. Large-basket weekly-shop operations and small-basket convenience operations therefore need different tolerance settings even at the same underlying rate.

Can route optimization reduce substitution rates? No, and it should not be bought on that basis. Research summarized by ECR Retail Loss attributes 38% of out-of-stock causes to store stocking, 22% to store forecasting and 9% to supply. Routing sits in the smallest slice. Its job is to produce plans that survive substitution, not to prevent it.