---
title: "Last-Mile Delivery Efficiency in 2026: How a Network-Wide Number Can Rise While Every Segment Gets Worse"
id: "27161"
type: "post"
slug: "last-mile-delivery-efficiency-network-mix-shift-2026"
published_at: "2026-09-28T15:00:00+00:00"
modified_at: "2026-09-28T16:15:10+00:00"
url: "https://locus.sh/blogs/last-mile-delivery-efficiency-network-mix-shift-2026/"
markdown_url: "https://locus.sh/blogs/last-mile-delivery-efficiency-network-mix-shift-2026.md"
excerpt: "Two delivery segments both improved utilization, one by 12%, one by 15%, yet the network-wide average fell 6.5% once volume mix shifted between them. The aggregate number and the underlying performance told opposite stories."
taxonomy_category:
  - "General"
---

#### [General](https://locus.sh/blogs/category/general/)

# Last-Mile Delivery Efficiency in 2026: How a Network-Wide Number Can Rise While Every Segment Gets Worse

[Aseem Sinha](/author/aseem_locus/)

Sep 28, 2026

16 mins read

Last-mile delivery efficiency is almost always reported as one network-wide number, a blended stops-per-vehicle-hour or cost-per-delivery figure covering every segment an operation serves. That single number can move in a direction no individual segment moved in, whenever the mix of volume between segments shifts at the same time performance within each segment changes. Modeling two segments with real efficiency gaps, both genuinely improving period over period, found the network-wide average falling once the volume share of the lower-efficiency segment crossed a specific, computable threshold, even though both segments were doing better than before. Locus, the world’s first Decision-Intelligent, Agentic TMS, reports efficiency by segment as well as in aggregate, so a mix shift is visible as a mix shift rather than mistaken for a performance decline or a performance gain that is not really there.

## Key Takeaways

- Two segments modeled from stated route archetypes ran at 3.97 and 2.18 stops per vehicle-hour respectively, and both improved genuinely, by 12% and 15%, period over period.
- With the lower-efficiency segment’s volume share fixed at 20% in both periods, the network-wide average rose 12.4%, correctly reflecting the improvement in both segments.
- Holding the same per-segment improvements but letting the lower-efficiency segment’s share rise to 55%, the network-wide average fell 6.5%, even though neither segment’s own efficiency went backward.
- The exact threshold where the network average flips from rising to falling is computable: in this model it sat at a 43% volume share for the lower-efficiency segment.
- Locus reports efficiency at the segment level as well as network-wide, so a shift in volume mix is distinguishable from a shift in performance rather than being blended into one ambiguous number.

## Why the Network-Wide Number Can Mislead: The Business Case

A blended efficiency figure answers a narrower question than it appears to. It tells you what happened to the average, not what happened to performance, and the two only match when the mix of work behind the average holds still. Most networks do not hold still, because the mix between dense urban delivery, suburban routes and sparse regional territory shifts with seasonality, new account wins, and channel growth, often on a faster cycle than the underlying routing performance changes.

This is a known statistical pattern outside logistics, most famously in the 1973 UC Berkeley admissions case, where [aggregate admission rates appeared to favor men over women applicants, while department-by-department data showed admission rates similar or even favoring women in most departments](https://www.statology.org/simpsons-paradox-when-aggregated-data-tells-a-different-story/)
. The paradox arose because departments preferred by women applicants happened to have lower admission rates for everyone, so the mix of applications across departments drove the aggregate figure more than any actual bias within a department did. The mechanism transfers directly: a network-wide efficiency number is an admissions-style aggregate across segments with structurally different efficiency ceilings, and the mix between them can dominate the number the same way department mix dominated the Berkeley figures.

The business version of this shows up whenever growth is uneven across segments. A company winning suburban and exurban accounts faster than urban ones, because that is where the addressable growth sits, is mechanically diluting its network-wide efficiency even while every individual segment, urban and suburban alike, is getting measurably better at what it does. Reading the blended number alone tells a management team the wrong story at exactly the moment they most need the right one, since that is usually also when they are deciding whether the routing program is working.

The stakes are proportional to how uneven the mix shift is and how far apart the segments sit to begin with, both of which are common in last-mile networks that span dense metro cores and much sparser surrounding territory within the same reporting unit.

There is also a governance cost to getting this wrong that compounds over time. A team whose real improvement gets read as a decline because of mix effects outside their control has little reason to keep improving, and a team whose stagnation gets read as progress because of a favorable mix shift has no signal telling them to change anything. Both effects push in the same direction: toward a routing program that stops earning the credit or the correction it actually deserves.

| Also Read: Last-Mile Delivery Efficiency: 2026 Complete Guide |
| --- |

## How a Mix Shift Overwhelms a Real Improvement

### 1. Two segments start with genuinely different efficiency levels

A dense urban segment and a sparse regional segment do not operate at the same stops-per-vehicle-hour, for reasons rooted in geography rather than management quality. This gap is structural, not a performance gap to be closed.

### 2. Each segment improves on its own terms, period over period

Better sequencing, better dispatch, fewer wasted miles, each segment gets measurably more efficient at what it does. Both improvements are real and both would look good reported in isolation.

### 3. The network reports one blended number across both

Total stops divided by total vehicle-hours across the whole network, which is the number that reaches a board slide or an investor update. It is a volume-weighted average of the two segment figures.

### 4. Volume mix between the segments shifts in the same period

New accounts, seasonal demand, or channel growth changes how much of total volume each segment now represents, independent of anything happening to routing performance.

### 5. The weighting shift can outrun the improvement

If the lower-efficiency segment’s volume share rises enough, its lower number pulls harder on the average than either segment’s own improvement pushes it up. The arithmetic works against the true story once the weighting moves past a threshold.

### 6. The reported average moves opposite to both underlying trends

The network-wide figure falls even though every segment it is built from improved. Read alone, it looks like the routing program failed. Read by segment, it did not.

| Also Read: Regional Last-Mile Delivery Efficiency in US Cities |
| --- |

## What the Model Shows

The model uses two route archetypes built from stated inputs rather than observed customer data: a dense urban segment and a sparse regional segment, each planned with balanced angular clustering, nearest-neighbor sequencing and local search. Baseline utilization for the two segments came out at 3.97 stops per vehicle-hour for urban and 2.18 for regional, a gap consistent with the density relationship modeled elsewhere in this cluster. Each segment is then given a genuine improvement, 12% for urban and 15% for regional, reflecting real gains a routing program might plausibly deliver to each. The regional segment’s volume share is then varied across two periods to see what the network-wide weighted average does.

**With the mix held stable, the average correctly reflects the improvement.** At a 20% regional volume share in both periods, the network average rose 12.4%, tracking the genuine gains in both segments. This is the case where the blended number tells the truth.

**As regional’s volume share rises even modestly, the picture starts to blur.** At a share rising from 20% to 30%, the blended average grew only 7.0%, materially understating the real improvement happening inside both segments. The mix shift was already eating into the reported gain well before it reversed it.

**Past a share of roughly 40%, the mix shift outweighs the improvement entirely.** At a share rising to 55%, the network-wide average fell 6.5%. At a share rising to 70%, it fell 14.5%. In both cases, urban improved 12% and regional improved 15%, exactly as in the first scenario where the average rose 12.4%. Only the mix changed, and the direction of the headline number flipped.

**The threshold is a real, computable number, not an approximation.** Solving for the exact volume share at which the network average stops rising and starts falling put it at 43% for this model’s specific efficiency gap and improvement rates. Below that share, the reported average understates or correctly reflects the improvement. Above it, the average actively misreports the direction of underlying performance.

**What the model does not settle.** It uses one pair of segments and one pair of improvement rates; a network with a wider or narrower efficiency gap between segments, or with more than two segments, will have a different threshold, though the same mechanism. It also assumes the improvement rates are genuinely independent of the mix shift, when in practice a segment absorbing a lot of new volume quickly may see its own improvement rate affected by that growth, which this model holds separate for clarity. A network with three or more segments compounds the effect further, since each pair of segments has its own threshold and the overall aggregate is subject to whichever combination of mix shifts happens to be moving at any given time, which is harder to reason about intuitively but no different in kind from the two-segment case modeled here.

| Also Read: Delivery Performance KPIs for 2026 |
| --- |

## Segment-Level and Network-Wide Reporting: Key Differences

| Dimension | Network-wide average only | Segment-level plus network-wide |
| --- | --- | --- |
| What moved is visible | No, mix shift and performance change are combined into one number | Yes, each is isolated and separately attributable |
| Risk when mix is stable | Low, the average is a fair summary | Low, same information either way |
| Risk when mix is shifting | High, the average can move opposite to actual performance | Low, segment figures show the true trend regardless of mix |
| Diagnosing a falling average | Cannot distinguish a real decline from a mix shift | Immediate, the segment figures either confirm or rule out decline |
| Reporting cadence needed | Whatever the network-wide figure is tracked on | Same cadence, segment cut added at no extra collection cost |
| Typical failure | A working routing program is judged to have failed | None, the program’s actual effect is visible |

## What to Look for in Segment-Aware Efficiency Reporting

### Efficiency reported by segment as the default, not an occasional drill-down

If segment-level figures only get pulled when the network-wide number looks wrong, the mix-shift problem has already had time to mislead a decision before anyone checked. The segment cut should be standing, not reactive.

### Volume mix tracked as its own metric alongside efficiency

The share of total volume each segment represents, period over period, is what determines whether the network-wide average is a fair summary or a misleading one. It deserves its own line on the same report.

### The mix-shift threshold computed for the network’s own segments

The volume share at which a mix shift can overwhelm real improvement is calculable from a network’s own segment efficiency gap. Knowing that number in advance tells a team how much mix movement its reporting can tolerate before the aggregate stops being trustworthy.

### Period-over-period comparisons decomposed into mix effect and performance effect

Rather than reporting a single percentage change in the blended number, the change should be split into the portion caused by segments individually improving and the portion caused by volume reweighting between them. Both are real; only one describes routing performance.

### Alerts triggered on segment divergence from the network trend, not just on the network trend itself

A segment whose own efficiency is falling should surface on its own terms, rather than being masked by other segments moving the aggregate in the opposite direction. The network-wide number is not a substitute for segment-level monitoring, only a summary of it.

| Also Read: What Is Fleet Utilization? Key Metrics & Importance |
| --- |

## Segment Reporting in Practice

**A Fortune 50 parcel and logistics network.** More than a million freight shipments a year across 51 sites and a 4,500-strong driver pool, with each site’s performance previously rolled into network figures that obscured which sites were actually improving. Centralizing raised weekly execution from 75% to 92% and surfaced more than $14M in unused capacity, including $565K identified at a single site, a result only visible once performance was measured site by site rather than as one blended network figure.

**A leading North American retailer.** Ocean, rail and road ran through six separate legacy systems, each reporting on its own basis, so a network-wide efficiency figure could not have been trusted even if one had existed. Consolidation onto a single system produced more than $1M in savings with 99%+ on-time store delivery and 95%+ route compliance, made possible by comparable, segment-level reporting across every mode rather than an opaque blended total.

**A global FMCG distribution network.** Ten Asian countries, 1,000+ distributors and 5,000+ riders, spanning wildly different densities and route archetypes within one operation. 12,000+ trips a month were eliminated against $4B+ in optimized orders at 3X ROI, a result that required understanding which specific parts of the network were improving and by how much, since a single blended figure across ten countries of varying density would have hidden exactly where the gains were coming from.

## Common Mistakes in Reading Network-Wide Efficiency Numbers

**Treating a falling network average as proof the routing program failed.** It can just as easily reflect volume shifting toward a structurally lower-efficiency segment while every segment individually improves. The segment-level breakdown is the only way to tell the two apart.

**Treating a rising network average as proof everything is improving.** The reverse mix shift can flatter a network-wide number even while a segment’s own performance is flat or declining, if that segment’s volume share is shrinking at the same time. A rising average deserves the same segment check as a falling one.

**Comparing network-wide figures across periods without checking mix stability first.** Any period-over-period comparison of a blended metric implicitly assumes the mix behind it held still. Where it did not, the comparison is measuring mix change as much as performance change, without saying so.

**Setting a single network-wide efficiency target across segments with very different structural ceilings.** A regional segment held to an urban efficiency target will always look like it is failing, and an urban segment held to a blended target will always look like it is coasting, regardless of how well either is actually managed.

| Also Read: Top 10 Last-Mile Delivery Metrics to Track in 2026 |
| --- |

## How Locus Reports Efficiency Without the Mix-Shift Blind Spot

Locus, the world’s first Decision-Intelligent, Agentic TMS, reports efficiency at the segment level as a standing default rather than a drill-down reserved for when the network-wide figure looks wrong, which is what keeps a mix shift from being mistaken for a performance change or a performance change from being masked by a mix shift. The [route planning and dispatch layer](https://locus.sh/route-planning-system/)
 reasons across more than 250 real-world operating constraints for every segment it plans, so the underlying route data needed to separate mix effects from performance effects already exists rather than needing to be reconstructed after the fact. Six governance mechanisms covering explainability, traceability, evaluation, autonomy levels, execution sandbox and human-in-the-loop keep every efficiency figure traceable to the routes and constraints behind it, at whatever level of aggregation it is viewed. The [Control Tower](https://locus.sh/control-tower-software/)
 carries the executed record for every segment, so a period-over-period comparison can be decomposed into the portion driven by volume reweighting and the portion driven by genuine routing performance, rather than delivered as one ambiguous blended percentage.

The platform reasons across those constraints over 1.5B+ deliveries for 360+ enterprise customers in 30+ countries at 99.99% uptime, with $320M+ in aggregate logistics cost savings, 800M+ miles reduced and 17M+ kg of CO2 avoided. Locus has been [recognized by Gartner for seven consecutive years](https://locus.sh/analyst-recognition/)
, including the 2026 Gartner Hype Cycle for Supply Chain Execution and Logistics Technologies and the 2026 Gartner Market Guide for Multicarrier Parcel Management Solutions, where ShipFlex is featured as a Representative Vendor. Locus holds Leader designation in the QKS SPARK Matrix for Transportation Management Systems 2025 and the #1 position for Route Planning in 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.

Two deployments show what segment-level visibility caught that a blended figure would have missed. A [Fortune 50 parcel and logistics network](https://locus.sh/case-studies/fortune-50-parcel-centralized-dispatch/)
 running more than a million freight shipments a year across 51 sites lifted weekly execution from 75% to 92% and exposed more than $14M in unused capacity, including $565K found at a single site, once each site was measured on its own terms rather than folded into a network average that would have hidden exactly where the opportunity sat. A [global FMCG distribution network](https://locus.sh/case-studies/global-fmcg-logistics-automation/)
 running across ten Asian countries with 1,000+ distributors and 5,000+ riders eliminated 12,000+ trips a month against $4B+ in optimized orders at 3X ROI, a result that depended on knowing which parts of a structurally uneven network were actually responding to the routing program rather than trusting one number to speak for all of them.

A network-wide efficiency figure is a volume-weighted average, and like any weighted average it can move opposite to every component it is built from whenever the weights shift. Modeling two segments with a real efficiency gap, both genuinely improving, found the network-wide average rising 12.4% with volume mix held stable and falling 6.5% with the same improvements but a mix shift toward the lower-efficiency segment, crossing from one direction to the other at a computable 43% volume-share threshold. A routing program working exactly as intended can be read as failing, or a stalled one can be read as succeeding, purely from how volume happened to redistribute in the same period. Locus reports efficiency by segment as a default, so a mix shift is visible as a mix shift. [Request a Locus segment efficiency review](https://locus.sh/schedule-demo/)
 to see what your own blended numbers might be hiding.

## Frequently Asked Questions

**Can a network-wide efficiency number fall even if every segment is improving?** Yes. Modeling this directly found the network-wide average falling 6.5% while both underlying segments improved, 12% and 15% respectively, because the lower-efficiency segment’s volume share rose enough to outweigh the improvement in the weighted average. The mechanism is a volume-weighted average responding to a change in weights, not to a change in the numbers being weighted.

**What is Simpson’s paradox and how does it apply to delivery efficiency?** Simpson’s paradox is the statistical pattern where a trend visible in every subgroup of data reverses or disappears once the subgroups are combined, because the mix between subgroups is itself doing work on the aggregate figure. It applies directly to a network-wide delivery efficiency number, which is a weighted average across segments with different structural efficiency ceilings.

**At what point does a volume mix shift start overwhelming real improvement?** It depends on the efficiency gap between segments and the size of each segment’s improvement, but it is computable for any specific network. In this model, with a roughly 1.8-times efficiency gap between segments and improvements of 12% and 15%, the threshold sat at a 43% volume share for the lower-efficiency segment.

**How do you tell whether a change in a blended efficiency number is real?** Break the period-over-period change into the portion caused by each segment’s own performance and the portion caused by volume reweighting between segments. If the segment-level figures are moving in the same direction as the blended number, the blended number is trustworthy. If they diverge, the mix shift is doing the work.

**Should efficiency targets be set at the network level or the segment level?** At the segment level, because segments with genuinely different structural efficiency ceilings cannot be held to one fair target. A network-wide target either sets an unreachable bar for sparser segments or an unambitious one for denser segments, and neither tells a segment manager anything actionable about their own performance.

**How often should segment-level efficiency be reported alongside the network-wide figure?** On the same cadence as the network-wide number, not as an occasional investigation. The segment cut uses data the network-wide figure is already built from, so there is no additional collection cost, and reporting it only when something looks wrong means the mix shift has already had time to mislead a decision.

MEET THE AUTHOR

Aseem Sinha

Vice President - Marketing

Aseem, leads Marketing at Locus. He has more than two decades of experience in executing global brand, product, and growth marketing strategies across the US, Europe, SEA, MEA, and India.

### Related Tags:

[https://locus.sh/blogs/last-mile-delivery-cost-density-benchmark-2026/](https://locus.sh/blogs/last-mile-delivery-cost-density-benchmark-2026/)
#### [General](https://locus.sh/blogs/category/general/)

## [Last-Mile Delivery Cost in 2026: What a Density Benchmark Actually Transfers](https://locus.sh/blogs/last-mile-delivery-cost-density-benchmark-2026/)

[Ishan Bhattacharya](https://locus.sh/blogs/author/ishan_locus/)

Sep 28, 2026

A 20-fold drop in stop density raised cost per delivery by only 1.69 times in this model, not the multiples often quoted. Importing a benchmark from a denser network still misprices work, just not in the direction usually assumed.

[Read more](https://locus.sh/blogs/last-mile-delivery-cost-density-benchmark-2026/)

[https://locus.sh/blogs/first-attempt-rate-roi-density-last-mile-2026/](https://locus.sh/blogs/first-attempt-rate-roi-density-last-mile-2026/)
#### [General](https://locus.sh/blogs/category/general/)

## [Last-Mile Delivery Efficiency and Cost in 2026: What a First-Attempt-Rate Improvement is Actually Worth, by Density](https://locus.sh/blogs/first-attempt-rate-roi-density-last-mile-2026/)

[Anas T](https://locus.sh/blogs/author/anas_locus/)

Sep 28, 2026

A one-point gain in first-attempt delivery rate is worth 2.2 times more in a sparse regional network than in a dense urban one. Applying one improvement target network-wide misallocates the investment behind it.

[Read more](https://locus.sh/blogs/first-attempt-rate-roi-density-last-mile-2026/)

## Last-Mile Delivery Efficiency in 2026: How a Network-Wide Number Can Rise While Every Segment Gets Worse

- Share
- [Print](javascript:window.print())
- [Download](#)
- [Schedule a Demo](https://locus.sh/schedule-demo/)

### Is your team spending more time on fixing logistics plan than running the operation?

- Agentic transportation management from order intake to freight settlement
- Route optimization built on 250+ real-world constraints
- AI-driven dispatch with automatic execution handling

20%Cost Reduction

66%Faster Planning Cycles

[Schedule a demo](/schedule-demo/)

Insights Worth Your Time

#### [General](https://locus.sh/blogs/category/general/)

## [Locus 2026 UK Consumer Survey: Why Returns Visibility is Now the Conversion Engine for AI-Driven Shopping in UK Retail](https://locus.sh/blogs/returns-visibility-conversion-engine-ai-shopping-uk-retail-locus-q2-2026-consumer-survey/)

[Aseem Sinha](https://locus.sh/blogs/author/aseem_locus/)

May 29, 2026

#### [General](https://locus.sh/blogs/category/general/)

## [Locus 2026 US Consumer Survey: Generative AI isn’t Just Changing How Consumers Shop, it’s Breaking the Demand Patterns US Retail Was Built On](https://locus.sh/blogs/generative-ai-shopping-effect-retail-fulfillment-operations-locus-q2-2026-consumer-survey/)

[Ishan Bhattacharya](https://locus.sh/blogs/author/ishan_locus/)

May 29, 2026

#### [General](https://locus.sh/blogs/category/general/)

## [Embedded vs Bolted-On AI: The Architecture Question European Logistics Buyers Are Asking](https://locus.sh/blogs/embedded-vs-bolted-on-ai-european-logistics-platform-architecture-business-benefits/)

[Aseem Sinha](https://locus.sh/blogs/author/aseem_locus/)

May 21, 2026

#### [General](https://locus.sh/blogs/category/general/)

## [Hybrid Fleet Management: How Owned, 3PL, Gig, ICE, and EV Capacity Actually Operate at Most Enterprises](https://locus.sh/blogs/three-workforce-fleet-reality-owned-3pl-gig-drivers/)

[Aseem Sinha](https://locus.sh/blogs/author/aseem_locus/)

May 7, 2026

#### [General](https://locus.sh/blogs/category/general/)

## [US Returns Hit $850 Billion in 2025: Why US Retailers Are Restructuring Reverse Logistics in 2026](https://locus.sh/blogs/850-billion-us-returns-ai-routing-reverse-logistics-2026/)

[Ishan Bhattacharya](https://locus.sh/blogs/author/ishan_locus/)

May 7, 2026