Physical intelligence. Enhanced by AI.

Financial reporting captures revenue. Satellite observations capture physical activity. These are related—but not identical.

Retail sales can increase because of higher prices, inflation, larger basket sizes and product mix changes, even when customer traffic is unchanged or declining.

Conversely, increasing physical activity may indicate improving customer demand before revenue growth becomes visible.

Axfood provides an ideal environment for demonstrating the methodology. The company operates several major grocery formats, including Willys, City Gross, Hemköp and Snabbgross.

This initial study focuses on Willys and City Gross because they provide large-format consumer retail locations with observable parking activity.

Coverage includes 186 monitored stores, approximately 80% of eligible locations and approximately 70% of Axfood consumer retail sales. Stores are excluded only when reliable measurement is not possible. No missing stores are estimated.

Each store is measured independently against its own historical baseline. The model combines three satellite-derived observation layers.

Sentinel-1 radar provides frequent observations regardless of cloud cover, darkness and weather conditions.

Sentinel-2 optical imagery provides additional information from surface characteristics and environmental conditions.

High-resolution parking activity detects vehicles from aerial imagery and is used as a direct physical indicator of customer presence. Counts are measured inside validated parking areas for each store—not the surrounding neighbourhood.

The same methodology produces a Store Activity Index (0–100) for every monitored location—designed for operators and asset owners who already control sales data but want an independent traffic layer.

A major challenge in retail analytics is that stores differ significantly. A large hypermarket naturally has more cars than a small urban store.

Therefore, the model does not compare absolute parking counts between locations. Instead, every store is measured against its own historical behaviour.

This removes much of the impact from store size, parking capacity, location and permanent format differences.

The model was evaluated out-of-time: developed on 2022 Q1 – 2024 Q4, tested on 2025 Q1 – 2026 Q2 without access to future data.

This is not millimetre-level forecasting. The signal is designed to show direction—whether physical activity at a store or across the chain is strengthening or weakening relative to its own history.

Backtests show the activity layer is consistently informative on that question. When traffic moves clearly, the signal and reported sales trends have often pointed the same way—particularly once price-driven periods are understood separately. In flat or inflation-heavy quarters, the two naturally diverge, which is why the product is positioned as an independent activity view rather than a sales substitute.

For retail operators who already control sales and LFL data, the product focus is an independent physical activity layer at store level.

Each monitored location receives a quarterly Activity Index from 0 to 100, measuring whether footfall is rising or falling relative to that store's own history. For stores in competitive catchments, we also report local market share—activity growth versus nearby grocery peers in the same quarter.

The index does not estimate sales. It measures traffic—designed to complement your internal reporting, not replace it.

During development, earlier versions produced substantially higher correlations. Additional testing showed that much of this performance was caused by unintended information leakage: calendar variables identified observation years, raw vehicle counts captured store size and image metadata introduced hidden shortcuts.

These effects were removed. After correction, performance decreased but remained statistically significant. We consider this improvement in reliability more important than maximizing headline correlation.

The production model reports only leakage-corrected results.

Ahead of Axfood's Q2 2026 earnings release (15 July 2026, 07:00 CET), the satellite activity signal offers an independent read on physical store behaviour — registered before the report, to be assessed against actual results.

This section presents the full quantitative analysis: the locked chain signal, a full-panel recalculation, an absolute-level comparison against historical Q2 norms, and a satellite-informed LFL estimate incorporating the macro backdrop.

The dataset measures physical activity. It does not directly measure kronor of sales, basket size, profitability or online grocery activity.

Online fulfilment represents a known limitation. Home delivery and click-and-collect generate revenue without necessarily creating parking activity.

This is why the dataset is designed as a complementary information source rather than a replacement for financial reporting.