UQO Working Papers · WP38 · v1.0
The Economic Geography of Digital Retail Supply: Product Diversity, Regional Specialization, and the Hidden Structure of Quebec's Online Economy
Evidence from 390,108 product listings of 1,569 Quebec online stores across 17 administrative regions (Fabri-Ka snapshot, 1 November 2026)
- Year
- 2026
- First release
- November 2026
- Latest revision
- November 2026
- Publication status
- Working paper
Regional specialization in Quebec's online retail is a property of merchants, not of listing volumes
On 390,108 listings of 1,569 stores, the region–category dependence beats 2,000 merchant reshuffles under merchant-equal weights (Cramér's V 0.112 vs 0.086) but is indistinguishable from chance under listing weights; catalogues are so concentrated (Gini 0.85) that the preregistered joint criterion fails and the complexity index is withheld.
Abstract
Can product-level online catalogues reveal the spatial organisation, diversity and specialization of a regional retail economy beyond what business counts show? This paper studies one read-only snapshot (1 November 2026) of Fabri-Ka, an aggregator of the public catalogues of 1,569 Quebec-based online stores (390,108 published listings, 17 administrative regions), with a preregistered design in which every regional statistic is confronted with a null that reallocates merchants rather than listings, and with three adversarial referee rounds whose corrections are reported.
After cleaning (370,054 listings, 1,557 stores) and recovering merchant locations from official geography, regional specialization turns out to be a property of merchants, not of listing volumes: under merchant-equal weights the region–category dependence exceeds 2,000 merchant reshuffles (Cramér's V 0.112 vs. 0.086, p < 0.001) in every sample, labelling and discovery-source variant, at the level of regions and of MRCs, whereas under listing weights it is indistinguishable from chance (p = 0.20; 0.43 after trimming the largest one per cent of stores), a low-power comparison because catalogue sizes are extremely skewed (Gini 0.85; the top decile of stores holds 79% of listings). By the preregistered joint criterion H1 is nevertheless not supported. Total regional variety rises with regional size, but the breadth of a typical set of fifteen merchants shows no detectable association with population, business counts, income or density across sixteen regions.
On 1,064 matched products sold by several merchants (matching precision 0.98–1.00), posted prices differ by a median of 4.7 log points; resellers post prices 11.6 log points below producers' own shops (p = 0.004), while the joint test of regional price levels does not reject equality once the wild cluster bootstrap is implemented correctly (p = 0.098) and rests on a single region. An ECI-type complexity indicator fails its preregistered bootstrap validation (median rank correlation 0.68 on a 17 × 20 matrix) and is withheld; the index family is published as separately validated components only. Measurement is central: the platform's keyword category labels reach 46% precision against annotator consensus and are replaced, by a rule fixed in advance, with a classifier validated at 88% on a held-out audit; a fifth of listings have no recoverable location; coverage per thousand business locations varies six-fold across regions. A local-origin price premium is untestable in this source. Code, audits, decision log, amendments and referee reports are released.
Key findings
390,108
listings of 1,569 Quebec online stores, 17 regions (370,054 / 1,557 after cleaning)
0.112 vs 0.086
Cramér's V of the region–category dependence against 2,000 merchant reshuffles (p < 0.001)
p = 0.20
the same dependence under listing weights: indistinguishable from chance
Gini 0.85
concentration of catalogue sizes; the top decile of stores holds 79% of listings
−11.6 log pts
reseller prices below producers' own shops on 1,064 matched products (p = 0.004)
46% → 88%
category-label precision: platform keywords vs the pre-specified replacement classifier
Scientific contributions
- A merchant-preserving null for regional statistics
- Every regional indicator is compared with a reallocation of merchants rather than listings, which separates the location of large catalogues from genuine specialization; under this null the merchant-level signal survives every variant and the listing-level signal vanishes.
- Preregistered verdicts, hypothesis by hypothesis
- H1 (specialization) not supported by its joint criterion, H2 (diversity and scale) inconclusive, H3 (concentration) supported, H4 (related diversification) exploratory, H5 split (channel supported, geography not), H6 untestable, H7 (complexity) not validated.
- Price dispersion across channels, not regions
- Resellers post 11.6 log points below producers on identical items while the joint test of regional price levels fails to reject equality once the wild cluster bootstrap is corrected and the influence of a single region is exposed.
- Measurement defects quantified and corrected
- Platform category labels (46% precision) replaced by a validated classifier (88%), a fifth of listings without recoverable location, six-fold variation in coverage per thousand business locations, and regional labels carrying a pre-2021 map, each measured and carried into the conclusions.
- Withdrawn first-draft findings reported as such
- A geographically structured price level and a power-law tail did not survive adversarial refereeing of the code and are reported as withdrawn, which is what the preregistered, auditable pipeline exists for.
Methodology
Methodology (5)
Preregistration with amendments
Seven hypotheses with joint decision criteria fixed before analysis; a decision log records amendments; three adversarial referee rounds over the code with corrections reported.
Merchant-reshuffling permutation tests
Region–category dependence (Cramér's V) under merchant-equal and listing weights against 2,000 reallocations of merchants, at region and MRC level, across sample, labelling and discovery-source variants.
Diversity, concentration and relatedness
Total regional variety versus rarefied breadth of a typical fifteen-merchant set; Gini and tail analysis of catalogue sizes; store-level relatedness density and category entry over ten weeks.
Price tests with wild cluster bootstrap
Channel (reseller vs producer) and regional price-level tests on matched products with correctly implemented wild cluster bootstrap and single-region influence checks.
Complexity indicator validation
ECI-type ranking subjected to a preregistered bootstrap over merchants on a 17 × 20 matrix; withheld when the median rank correlation (0.68) fails the threshold.
Research data
- Fabri-Ka catalogue snapshot
- Owner-authorised, read-only copy of the aggregator database taken with sqlite3 backup on 2026-11-01 (SHA-256 recorded): 390,108 published listings of 1,569 Quebec-based online stores; cleaned to 370,054 listings and 1,557 stores; not redistributed because catalogues belong to the merchants.
- Official geography and business counts
- Merchant locations recovered from official administrative geography (17 regions, MRCs); population, business counts, income and density per region from public statistics; discovery runs through regional directories, which drives the six-fold variation in coverage.
- Matched-product price sample
- 1,064 products sold by several merchants, matched with precision 0.98–1.00, used for the channel and regional price tests; an annotator-consensus audit of category labels and a held-out validation of the replacement classifier.
Reproducibility
- Deterministic pipeline (seed 20261101) from a single Makefile: ingest, clean, match, validate, features, models, robust, tables, figures, paper; make test runs invariant tests on the cleaned panel
- Raw snapshot not redistributed; data dictionary, sample-flow table, every experiment summary, decision log, amendments and referee reports released
- Category classifier, annotation audit and matching precision reported against held-out samples
- Full LaTeX source of the 51-page paper
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | November 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 51 p. |