1\section{Empirical methods and identification limits}\label{sec:methods}23\subsection{Weighting schemes}4Catalogue sizes are extremely skewed (Section~\ref{sec:results_h3}): the ten largest stores hold more than a third of all listings. Every regional statistic is therefore computed under three schemes. Under \textbf{L} each listing counts one. Under \textbf{M} (merchant-equal, the primary scheme for regional composition) each listing counts $1/n_s$, so that every store contributes one unit regardless of catalogue size. Under \textbf{T} listings count one after dropping the one per cent largest stores. The schemes answer different questions --- L describes what a visitor sees, M describes the population of merchants --- and we report where they disagree.56\subsection{Revealed assortment advantage and the merchant-reshuffle null}7For region $r$ and category $c$ let $N_{rc}$ be the (weighted) count of listings. The revealed assortment advantage is8\begin{equation}9\mathrm{RAA}_{rc}=\frac{N_{rc}/\sum_{c}N_{rc}}{\sum_{r}N_{rc}/\sum_{rc}N_{rc}},10\end{equation}11the Balassa index on listings. Unlike export RCA it does not measure production, trade or competitiveness; it measures what the merchants of a region publish relative to what Quebec merchants publish overall. Dependence between region and category is summarised by Cramér's $V$ of the $R\times C$ table and tested against a null that reshuffles region labels across \emph{stores} while keeping every store's catalogue and every region's number of stores fixed; the $p$-value is the share of 2{,}000 permutations with $V$ at least as large as observed. This is the dartboard logic of \citet{ellison1997} transposed to merchants: a region may look specialised simply because one large merchant is located there, and the null absorbs this. Uncertainty for each $\mathrm{RAA}_{rc}$ comes from 2{,}000 bootstrap resamples of stores; because resampling with replacement duplicates stores and biases $V$ upward, bootstrap intervals are reported for the cells and not for $V$, whose inference is by permutation.1213\subsection{Diversity and concentration}14Regional diversity is the normalised Shannon entropy of category shares, $\tilde H_r=-\sum_c p_{rc}\ln p_{rc}/\ln 20$, the effective number of categories $\exp(H_r)$ and the Herfindahl index, under L and M. Because entropy increases mechanically with the number of merchants, the primary H2 outcome is the \emph{rarefied} entropy: the mean over 2{,}000 draws of fifteen stores without replacement, which puts regions with 15 and 235 stores on the same footing; regions with fewer than fifteen stores are excluded from that statistic. Concentration is the Gini coefficient of listings across stores, top-percentile shares, the rank--size slope, and a discrete power-law fit with $x_{\min}$ chosen by KS minimisation compared with a log-normal on the same tail by a Vuong likelihood-ratio test \citep[following][]{ellison1997}.1516\subsection{Region--category network and complexity indicator}17Let $M_{rc}=\mathbf 1[\mathrm{RAA}_{rc}\ge1]$ under scheme M. Diversity is the row sum, ubiquity the column sum; the ECI is the second eigenvector of $\tilde M=D_r^{-1}MD_c^{-1}M^{\top}$ \citep{hidalgo2009,mealy2019}, with the sign fixed so that it correlates positively with diversity; the fitness index is the fixed point of the non-linear iteration of \citet{tacchella2012}. Category relatedness is the Hidalgo-type proximity $\phi_{cc'}=\min\{P(c\mid c'),P(c'\mid c)\}$ computed from co-occurrence of categories across \emph{stores}, and communities are obtained by Louvain over 200 seeds. Validation follows the preregistration: (V1) the Spearman correlation between the full-sample ECI and 500 store-bootstrap ECIs; (V2) the second eigenvalue of $\tilde M$ and the spectral radius of the bipartite adjacency matrix against 2{,}000 store-level region reshuffles \citep[a merchant-preserving analogue of the null models in][]{saracco2015}; (V3) the leave-one-region-out predictive correlation of the ECI with regional population, business locations and income; (V4) the normalised mutual information between community partitions across seeds.1819\subsection{Matched products and hedonic dispersion}20No GTIN exists in the source. Cross-merchant identity is inferred in tiers --- A: shared merchant SKU and token-set title similarity $\ge70$; B: same normalised brand and identical normalised title; C: same brand, same category and token-sort similarity $\ge90$; D: identical title without brand --- restricted to listings in the price sample and to pairs of different stores that are not mirror domains of one business. Tiers A and B form the primary sample and C a sensitivity; D is never used for inference. Precision is estimated on a labelled sample of 240 pairs (Section~\ref{sec:data}). For listing $i$ in match group $g$ sold by store $m$ in region $r(m)$,21\begin{equation}22\ln P_{i}=\alpha_{g}+\gamma_{r(m)}+\delta_{\mathrm{platform}(m)}+\beta_1\ln n_m+\beta_2\,\mathrm{Reseller}_m+\beta_3\,\mathrm{Certified}_m+\varepsilon_{i},23\label{eq:hedonic}24\end{equation}25estimated by within-group least squares with standard errors clustered by store. Because each store belongs to one region, a store fixed effect cannot be included alongside $\gamma_r$: region effects are identified only through match groups that span regions, which is why the sample size in \emph{stores} and the number of regions with at least five matched stores are reported, and why inference uses a wild cluster bootstrap (Rademacher weights by store, 9{,}999 draws, restricted residuals) for the joint test of $\gamma_r$ and for $\beta_2$. Prices are posted minimum-variant prices at the snapshot; shipping, taxes and transactions are not observed.2627\subsection{Dynamics}28The only temporal margin is listing entry and exit after store onboarding ended (24 August 2026). Weekly presence of each (store, category) pair is reconstructed from first and last observation dates; relatedness density is computed from the baseline weeks 36--39 and entry is measured over weeks 40--44, a temporal holdout that prevents the outcome from feeding the regressor. A linear probability model with store and category fixed effects and store-clustered standard errors is compared with a placebo that relabels categories in the proximity matrix. The exercise is exploratory by preregistration.2930\subsection{External data and the representativeness audit}31Regional covariates come from the ISQ (population at 1 July 2025, business locations with employees in December 2025 by NAICS, disposable income per capita 2024, regional GDP 2024) and from Statistics Canada (Census 2021 land area and density on 2021 economic-region boundaries); vintages and concordance issues are recorded in the external-data manifest. Coverage is the number of Fabri-Ka stores per thousand business locations and per 100{,}000 inhabitants; the audit also tabulates discovery source and platform by region and category mix by platform, following the data-defect logic of \citet{meng2018}.32