1\section{Economic interpretation, policy relevance and limitations}\label{sec:discussion}23\subsection{What the online catalogue of a region is}4The results are consistent with a simple reading. Quebec's online retail supply, as seen through an aggregator of independent merchants, is overwhelmingly a population of small catalogues (median 35 listings) dominated in volume by a handful of very large ones. Regional specialization exists at the level of \emph{who} sells online --- maple and alcohol producers in Estrie, fashion and jewellery designers in Montréal, pet-supply and alcohol merchants in Montérégie --- but once listing volumes are counted, a region's assortment is largely the catalogue of whichever large merchants happen to be located there, and the data cannot tell this apart from a random relocation of merchants: the listing-weighted null is wide, so this is a statement about power as much as about the world. This distinction between merchant-level and listing-level specialization is the paper's main descriptive finding and a caution for any use of catalogue counts as regional indicators. Diversity rises with the number of located merchants, as business-type diversity rises with city size \citep{youn2016,hong2020}; measured on an equal footing of fifteen merchants it is narrow in range (0.73--0.89 on a 0--1 scale) and shows no detectable association with regional scale, although a sample of sixteen regions can only detect large associations. Posted prices of the same product differ across merchants by a median of about five log points and half of the matched groups differ by more than five per cent; the dispersion is structured by channel --- resellers and collective storefronts post lower prices than the producers' own shops for the same item, which is the opposite of a retail margin and is consistent with producers using their own shop as a list-price channel while discounting through distributors --- while its geographic component rests on one region and does not survive a correct bootstrap. The mechanism cannot be identified here.56\subsection{Policy relevance}7Three uses follow, each within the limits of the data. First, for ``buy local'' and regional branding policies, the certification registries and regional agri-food brands that feed the aggregator are themselves the main determinant of which merchants are visible online; a region's online presence is partly a function of its directories. Second, the ECI-type complexity indicators that are increasingly used in regional policy \citep{balland2019} do not survive bootstrap validation on a 17$\times$20 matrix of online categories: a policy indicator built this way would reshuffle regional rankings from one resample of merchants to the next, and we therefore withhold both the complexity component and a composite score. Third, the within-product price dispersion across merchants is large enough to matter for consumers comparing local producers with distributors, and the collection plan in Appendix~\ref{app:prereg} describes how a repeated snapshot with store pinning would turn the cross-section into a price panel.89\subsection{Limitations}10The limitations ledger in the repository lists fifteen items; the five that bound every conclusion are the following. (i) Listings are posted offers of located merchants: nothing here measures production, sales, employment or product origin. (ii) The discovery process is non-probability and region-correlated: coverage per thousand business locations varies six-fold, and 19\% of the assortment sample has no recoverable location. (iii) Category labels are measured with error: the platform's keyword rules reach 46\% precision, the classifier we substitute is validated on an independent audit but trained on model annotations, and all category results are reported under both labellings. (iv) With sixteen or seventeen regions the scale association and the complexity validation have little power, and the region effects in the price regression are identified only through products sold in several regions and dominated by one reseller. (v) The snapshot is a single cross-section without price history; dynamics are confined to ten weeks of listing entry and are exploratory.11