One checkout, two kinds of territory
Trade between China and Russia has moved a large share of consumer orders onto platforms. The published work on that trade spends most of its space on customs clearance, certification, tax rebates and payments [12]. Those frictions are real. They are also not what the buyer renegotiates on the next order. Once the parcel is in the country, the decision that repeats is where to collect it, and how far that place is from the usual walk home.
Reviews of cross-border e-commerce say little about this step. Hazarika and Mousavi (2022) find the field concentrated on China and the European Union, with logistics among the thinner topics. Even the better-covered flow of parcels into China treats the design of the distribution network as unfinished business [9]. The Russia-bound last mile, and the collection point in particular, sits further out on that margin. AliExpress Russia is a useful place to look, because one interface sits on top of a country that does not have one delivery geography.
Most buyers, and most of the dense street network, are in the European part of the country. Siberia and the Far East are large, sparsely settled, and expensive to cover with the same spacing of doors. Inside a single city the contrast is smaller but the same in kind. In Moscow or St. Petersburg a user is often a few hundred metres from several pick-up points, sometimes run by different firms. On the edge of those cities, and in a lot of provincial centres, one operator is the realistic choice, and “nearby” on the map can still mean a bus. Pedestrian crossings, the condition of the pavement and whether the point sits on a route people already walk all vary with the district. The platform does not experience this as five products. The user does.
Last-mile work already frames the design choice as push or pull: bring the parcel to the door, or bring the customer to a point [14]. Russian marketplaces mostly pull. Collection points are an old answer to that choice. Dutch evidence is careful about how strong the pull is. Shoppers were more willing to use a service point when several sat close to home, and a drive of about five minutes already looked like the practical limit [17]. Intention to use self-collection, in another setting, depends on whether the service fits existing habits and has a clear advantage over home delivery, not on age or employment [19]. Fulfillment method also shows up in repurchase intentions on its own, apart from the goods [4]. A count of points, with no route attached, is a weak description of any of this.
What the walk does to the number of orders
In the AliExpress Russia data the association is visible operator by operator. For each network the correlation is between the user’s walking distance to the nearest point and the number of orders that user placed. Pearson’s r is negative throughout: −0.41 for Cainiao, −0.49 for KazanExpress, −0.53 for X5, −0.63 for Russian Post and −0.67 for CDEK. On the pooled sample it is −0.58. Mean walking distances run from about 560 m (Cainiao) to about 830 m (CDEK). Mean orders per user are higher on the shorter networks (6.1 for Cainiao, 5.8 for KazanExpress, 5.2 for X5) and lower for Russian Post (3.9) and CDEK (4.1). The busiest band is roughly 200–500 m. Past that, frequency falls. If another operator has a closer door, some of the fall is switching rather than a lost order.
Table 1
Walking distance, orders per user, correlation, and the slope of orders on distance
|
Operator |
Mean distance, m |
Mean orders |
r |
Slope per km |
|
Cainiao |
560 |
6.1 |
−0.41 |
−2.218 |
|
X5 |
650 |
5.2 |
−0.53 |
−2.732 |
|
KazanExpress |
780 |
5.8 |
−0.49 |
−2.491 |
|
Russian Post |
725 |
3.9 |
−0.63 |
−3.604 |
|
CDEK |
830 |
4.1 |
−0.67 |
−3.851 |
|
Pooled sample |
— |
— |
−0.58 |
−2.945 |
Note. Means and correlations are the operator-level figures. Slopes are the distance coefficients from linear regressions of order counts on walking distance in kilometres. The pooled slope is estimated on 52,480 observations (standard error 0.018).
Read off the pooled slope, an extra kilometre of walking goes with just under three fewer orders in the window the data cover. The intercept, 6.81, is the fitted count at a zero distance. It is an anchor for the line, not a claim that anyone collects a parcel at the doorstep. The correlation of −0.58 means this one variable accounts for about a third of the variation in orders per user (r² ≈ 0.34). That matches the reported t-statistic on the same regression. It is still one variable. Region, income, and how long the user has been attached to a particular door are not in the equation, so the third is an upper reading of what distance does by itself.
The slope is not common property of the five networks. Cainiao’s coefficient is −2.22 orders per kilometre, X5’s is −2.73, KazanExpress −2.49, Russian Post −3.60 and CDEK −3.85. A separate description, not the regression line, has Cainiao’s average falling from about 6.6 to 5.4 orders when the route is a kilometre longer, and Russian Post’s from about 4.6 to 3.5. The drops are about 18 % and about a quarter because the starting levels differ. They should not be quoted as if they were the regression coefficients. The direction is the same, and the operators who work the thinner territories have the steeper lines.
It is easy to call the steeper line an elasticity. The sign would survive the relabelling. The estimated object would not. What was calculated is a slope in orders per kilometre. An elasticity is a ratio of percentage changes, and it was not the number reported. Where the networks overlap, the user can leave a bad door for another firm’s. Cainiao and X5, concentrated in denser urban fabric, lose orders more slowly as the walk grows, partly because a second point is often available. Russian Post and CDEK, with more of their volume in places where they are the only realistic option, do not offer that exit. A longer walk there is not a reason to switch. It is a reason to order less often. In the Far Eastern regions, where those two operators dominate, the extra distance behaves like a barrier rather than a mild inconvenience. Competition softens the distance effect only when a second door is actually there.
There is a plain cost reading of the pattern, and it should stay plain. Transaction-cost theory is written for firms deciding how to organise a deal: finding the other party, settling terms, enforcing them [18]. The buyer’s version in this setting is smaller and physical. Time spent walking to the point never appears on the receipt. Service-convenience research treats that time and effort as part of the service [2], and later work links convenience to whether the customer comes back [15]. The Russian coefficients fit that reading. They do not require a new label for it.
The regression cannot say the platform caused the orders. People who live near a convenient point may order more for reasons that have nothing to do with the pavement. Users also stick. A familiar door keeps getting selected even when a slightly better one opens nearby, and the advantage is small enough that switching is not worth the bother. A recommendation that only restates a choice the user would have made anyway will look tidy and change little. That limit is why the checkout test is in the paper. The correlations describe the constraint. They do not show that the constraint can be moved.
The line on the map and the path on the pavement
Geographical distance, the straight line between two coordinates, assumes the person can travel the line. Streets do not offer that. Accessibility is produced by the network, the crossings and the things that block them [8]. The gap between a straight line and travel along a real network is documented for access to other services, where the shortcut on the map is not the trip taken [3]. The pickup sample has the same kind of gap, and here it is large enough to change which point ranks first.
A straight line of 180 metres can become a walk of 850 metres if the short line crosses a highway, an industrial plot or a closed courtyard. Cases of 150–200 metres on the map and more than 800 metres on foot come up often enough, in the large cities in the sample, that they should not be filed as odd blocks. The usual obstacles are ordinary: a road with no nearby crossing, a railway, a fenced residential compound, an industrial territory, a river embankment, sometimes a change in elevation. The user does not experience “180 metres”. They experience the detour, the minutes, and whether the route is one they already use. Order frequency lines up with that walk more closely than with the straight line. Estimated on geographical distance alone, the same relationship leaves more of the variation in orders unexplained.
This is also why some well-placed pins are quiet. A point chosen because the rent was lower, or because it fell inside a zone drawn for couriers, can be a poor pedestrian destination. A point at a metro exit, beside a shop door, at the entrance of a residential block or on the way from a bus stop often takes more orders than a formally closer one. Two points with the same straight-line distance can produce different order rates if one of them sits across a highway. Zheng, Morimoto and Murayama (2020) locate pickup points in a Guangzhou district with a network Huff model for this reason: proximity measured through the street network is the quantity that demand responds to. From the operator’s side, travel imposed on the customer is lost demand and belongs inside the location calculation, not after it [5]. Adding points without regard to the pedestrian network risks a map that looks fuller and a set of doors people still do not use.
For a recommender the consequence is narrow. Sorting candidate points by straight-line distance keeps offering the door the user cannot reach without the detour. Sorting by walking distance will sometimes offer a point that looks worse on the map and is shorter on the pavement. The second rule is the one that was put in front of users.
Whether the screen can move a habit
People repeat a collection point. The experimental question is not whether logistics in general can be optimised. It is whether a change in how the point is shown changes the point that receives the order, the length of the walk, and the split of orders across operators, under conditions that are otherwise the same.
The design is a randomised comparison of the sort set out by Fisher (1935) and, for online products, by Kohavi, Tang and Xu (2020). Assignment was at user level. Each user who entered the test was placed in one group and stayed there, so the same person did not see two versions of the screen. The draw did not depend on region, order frequency, the operator used before, or other history. Three groups shared the catalogue, the prices, the promotions, the season and the set of points on the ground. Seasonality and one-off campaigns, if they hit, hit all three. The control group saw the ordinary selector, with no recommendation layered on it. In the soft treatment the usual point stayed pre-selected, and a more walkable alternative was shown beside it. Ignoring it cost nothing. In the hard treatment the algorithm’s point was already selected. Keeping another point meant changing the selection by hand.
The hard treatment is a default. People leave defaults in place more often than a fully deliberate choice would predict, even when opting out is a small action [11]. They also keep the current option when the gain from switching is not obvious [16]. If that carries over to a pickup screen, the pre-selected point should take a larger share than a suggestion sitting next to the old choice. It did. Both treatments lowered the average walking distance of the point users confirmed. The hard treatment lowered it by more. Order frequency rose with the shorter walk. That is the direction the regression implied, and the direction a cross-section on its own could not establish, because the cross-section never moved the distance while holding the user still.
The mix of operators moved as well. Share shifted toward networks whose points sit on shorter walks, Cainiao’s in particular, and away from operators that were the long walk in that district, including X5 or Russian Post where their local door was the inconvenient one. This is not a ranking of firms in the abstract. The same operator can gain in one city and lose in another, depending on where its doors actually are. Some of the shift outlived the first click. Users who accepted the recommended point tended to use it on later orders, including when the screen was no longer pushing it. A recommendation that matches the real walk can reset the habit. One that fights the walk will be declined, or accepted once and abandoned.
Two limits sit next to that result, and both are part of the result. The test compares presentations. It does not build points. In a settlement with a single usable door the algorithm has nothing better to offer, and a confident default would only hide the fact. A hard pre-selection also amplifies a bad route file. If the walking distances are wrong, the default sends more people to the wrong door, not fewer. The soft prompt moves less demand and does less damage when the network file is incomplete. What the test shows is the gain from ranking points that already exist, by the route the user will actually walk.
What follows for the platform, the operators and the thin regions
The coefficients and the experiment sit on three ideas that are already in the literature. None of them has to be renamed to fit.
Part of the cost of the order is paid in minutes at the end. Williamson (1979) was writing about the costs of arranging a transaction between firms. Collection effort is the consumer-side cousin of that idea, not a case he stated. The slope on walking distance is the trace in these data: a longer walk, fewer orders. Shortening the walk by changing the rank order is cheaper than opening points, and it only works where a shorter walk is already on offer.
A stock of pick-up points is not, by itself, difficult to match. Barney (1991) keeps the language of advantage for assets that are valuable and hard to imitate. Several operators in this sample already have national coverage, so the list of doors fails the test. What is harder to copy is the joint use of that list and a ranking built on pedestrian routes. The points were there before the test. The change in the treatment groups came from which door the screen put forward. That is the sense in which the algorithm and the network are one capability rather than two assets side by side.
Supply-chain integration is usually scored between firms: shared plans, shared information, a common operational picture [7]. The marketplace version is thinner than that literature. The platform does not merge Cainiao, X5, KazanExpress, Russian Post and CDEK. It does steer the buyer’s collection choice toward the operator who can serve that address with a shorter walk, and the steering shows up as a change in order share. Calling this integration is fair only if the word stays modest. It is coordination of the last choice in the chain. It is not a joint venture.
For the platform the working rule is specific. Rank by the walking route, not by the straight line, and try the default on a random subset before it becomes the only interface. A hard pre-selection moves more orders. It also punishes errors in the route data, so it should wait until those data are good in the cities where the detour problem is worst. For an operator, a point on a route people already walk is worth more than another pin in a district that already has a walkable door. The next budget does less if it raises the count of points and leaves the unwalkable ones open. Moving or closing the doors users avoid is the less comfortable decision, and it is the one the order data support. Rent and administrative boundaries are weak guides. They produce locations that look correct on a zoning map and stay quiet.
Policy splits along the same line as the sample. Harmonising customs, documents and payments between Russia and China addresses the cross-border leg [12]. It does not put a collection point in a town that lacks one, and it does not build a crossing over a road that currently forces an 800-metre detour. Low-density regions, and the Far Eastern districts where a single operator carries the network, remain an infrastructure problem. A recommendation is a tool for places where the infrastructure is already good enough to choose among. Treating it as a substitute for that infrastructure describes a country this sample does not contain.
The order data and the experiment support a limited claim. Walking distance predicts how often people order, and it does so through the route on the pavement rather than the line on the map. The checkout screen can change which distance they accept, when a better point is already there, and a pre-selected point changes it more than a prompt they can ignore. The screen does not close the gap between a courtyard with four doors and a town with one post office. That gap is still the main fact about last-mile logistics on this corridor.
References:
- Barney, J. (1991). Firm resources and sustained competitive advantage. Journal of Management, 17(1), 99–120.
- Berry, L. L., Seiders, K., & Grewal, D. (2002). Understanding service convenience. Journal of Marketing, 66(3), 1–17.
- Boscoe, F. P., Henry, K. A., & Zdeb, M. S. (2012). A nationwide comparison of driving distance versus straight-line distance to hospitals. The Professional Geographer, 64(2), 188–196.
- Boyer, K. K., & Hult, G. T. M. (2006). Customer behavioral intentions for online purchases: An examination of fulfillment method and customer experience level. Journal of Operations Management, 24(2), 124–147.
- Deutsch, Y., & Golany, B. (2018). A parcel locker network as a solution to the logistics last mile problem. International Journal of Production Research, 56(1–2), 251–261.
- Fisher, R. A. (1935). The design of experiments. Oliver & Boyd.
- Flynn, B. B., Huo, B., & Zhao, X. (2010). The impact of supply chain integration on performance: A contingency and configuration approach. Journal of Operations Management, 28(1), 58–71.
- Geurs, K. T., & van Wee, B. (2004). Accessibility evaluation of land-use and transport strategies: Review and research directions. Journal of Transport Geography, 12(2), 127–140.
- Giuffrida, M., Mangiaracina, R., Perego, A., & Tumino, A. (2017). Cross-border B2C e-commerce to Greater China and the role of logistics: A literature review. International Journal of Physical Distribution & Logistics Management, 47(9), 772–795.
- Hazarika, B. B., & Mousavi, R. (2022). Review of cross-border e-commerce and directions for future research. Journal of Global Information Management, 30(2), 1–18.
- Johnson, E. J., & Goldstein, D. (2003). Do defaults save lives? Science, 302(5649), 1338–1339.
- Karlovskaya, S., & Chelombitko, A. (2024). Prospects for the development of cross-border e-commerce between Russia and China. BRICS Journal of Economics, 5(2), 45–63.
- Kohavi, R., Tang, D., & Xu, Y. (2020). Trustworthy online controlled experiments: A practical guide to A/B testing. Cambridge University Press.
- Lim, S. F. W. T., Jin, X., & Srai, J. S. (2018). Consumer-driven e-commerce: A literature review, design framework, and research agenda on last-mile logistics models. International Journal of Physical Distribution & Logistics Management, 48(3), 308–332.
- Roy, S. K., Shekhar, V., Lassar, W. M., & Chen, T. (2018). Customer engagement behaviors: The role of service convenience, fairness and quality. Journal of Retailing and Consumer Services, 44, 293–304.
- Samuelson, W., & Zeckhauser, R. (1988). Status quo bias in decision making. Journal of Risk and Uncertainty, 1(1), 7–59.
- Weltevreden, J. W. J. (2008). B2c e-commerce logistics: The rise of collection-and-delivery points in the Netherlands. International Journal of Retail & Distribution Management, 36(8), 638–660.
- Williamson, O. E. (1979). Transaction-cost economics: The governance of contractual relations. The Journal of Law and Economics, 22(2), 233–261.
- Yuen, K. F., Wang, X., Ng, L. T. W., & Wong, Y. D. (2018). An investigation of customers’ intention to use self-collection services for last-mile delivery. Transport Policy, 66, 1–8.
- Zheng, Z., Morimoto, T., & Murayama, Y. (2020). Optimal location analysis of delivery parcel-pickup points using AHP and network Huff model: A case study of Shiweitang sub-district in Guangzhou City, China. ISPRS International Journal of Geo-Information, 9(4), 193.

