This post details an investigation into how Google Maps' ranking algorithm disproportionately allocates "survival" visibility to London restaurants, effectively acting as a market maker rather than a neutral directory. The author scraped over 13,000 London restaurants and developed a machine learning model (Gradient Boosting Regressor) to predict a restaurant's rating based only on structural features (e.g., review volume, price, location, chain status). This allows for the calculation of a "rating residual"—the difference between the actual rating and the algorithmically expected rating. Positive residuals indicate outperformance relative to algorithmic amplification, suggesting genuine value not currently captured by the platform's core signals (relevance, distance, prominence). The analysis extends to neighborhood level, clustering areas into "hubs" (elite, strong, everyday, weak) based on aggregated metrics, revealing that culinary diversity, particularly migrant cuisines, often clusters in areas with structurally weaker algorithmic visibility. The key takeaway is that platform ranking is a significant, opaque driver of local economic outcomes, suggesting a policy need for ranking algorithm transparency and auditability.
Investigating Algorithmic Power: Deconstructing Google Maps' Impact on London's Restaurant Economy
0:00 Motivation: Started by personal frustration with Google Maps recommendations; escalated into analyzing how digital platforms redistribute economic survival via algorithmic visibility.
Google Maps as a Market Maker: The platform uses signals like review volume, velocity, and brand recognition ("prominence") to rank listings, not just reflecting demand but actively organizing it.
Cumulative Advantage (Matthew Effect): Visibility generates reviews, which increases prominence, leading to more visibility. This disproportionately rewards established chains and high-footfall areas, creating a cold-start problem for new independents.
0:53 Building the Counterfactual Model: Developed a machine learning model (HistGradientBoostingRegressor) to predict a restaurant's rating based solely on structural features (cuisine, price, location, existing review count).
Cuisine Classification: Due to Google's poor/inconsistent cuisine tagging, a separate ML model was built to infer cuisine type from names and reviews for accurate competition analysis.
1:00 Rating Residual: Defined as Actual Rating - Predicted Rating. A positive residual signifies a restaurant performs better than the platform's visibility structure would suggest—a measure of algorithmic mispricing or overlooked value.
1:10 Caveat on Ads: Paid promotion visibility cannot be accounted for, meaning residuals might partly reflect undisclosed ad spend.
1:18 London Food Dashboard: The author built a dashboard (accessible at laurenleek.eu/food-map) allowing filtering for "underrated gems" (high positive residual).
2:00 Neighborhood Ecosystem Analysis: Restaurants aggregated into spatial hexagons; features computed (density, mean residual, chain share, etc.). PCA created a single "hub score" to rank ecosystem strength.
2:14 Hub Classification: Areas clustered into four types: elite, strong, everyday, and weak hubs, based on ecosystem strength metrics.
2:25 Cuisine Clustering: Culinary diversity is not evenly distributed; migrant cuisines often cluster in areas with weaker algorithmic visibility scores, tied to historical settlement patterns and affordability.
3:00 Policy Conclusion: Algorithmic invisibility can economically strand businesses regardless of local council efforts. Platform ranking systems with significant economic impact require mandatory transparency and auditability, analogous to auditing financial markets.
I don't care for the author's side-bar jokes about Jeremy Corbyn, but the core analysis regarding compounding advantage in digital ranking systems is sound. The methodology for isolating algorithmic effects via counterfactual prediction is reasonably robust for a first pass, assuming the feature set captures most non-quality related ranking drivers. The observation that platform visibility is now a crucial, un-audited lever of local economic policy seems highly probable (probability estimate: $\geq 0.9$).