On-Shelf Availability Is a Data Problem, Not a Logistics Problem

Introduction

On-shelf availability (OSA) is the measure of whether your product is on the shelf when the customer wants it. Simple in concept, brutal in practice. Most OSA initiatives fail—not because the logistics are wrong, but because the approach is. They're reactive: you discover stockouts when someone visits the store. Root causes are guessed, not known. Data from sell-out, inventory, and external factors lives in silos. The result is firefighting, not prevention.

We worked with a major European beverage distributor to flip that model. In 23 weeks, we deployed a demand forecasting engine with root cause classification and proactive alerting on GCP. The shift: from reactive discovery to data-driven prediction.

Why OSA Initiatives Fail

Reactive discovery — Field teams visit stores and find empty shelves. By then, the sale is lost. The customer has already chosen a competitor or left. Reacting to stockouts is too late.

Unknown root causes — Was it delayed restocking? Incorrect stock levels in the system? Poor shelf rotation? Promotional demand spike? Without classification, every stockout gets the same generic response. You can't fix what you don't understand.

Disconnected data — Sell-out data, inventory data, and external factors (promotions, weather, events) often live in different systems. No one has a unified view. Correlations are invisible. Predictions are impossible.

The Data-Driven Approach

We built three layers:

SKU/store level prediction — Demand forecasting at the granularity that matters. Each SKU in each store gets a 14-day horizon. We use sell-out history, inventory levels, promotions, and external signals. The model predicts where stockouts are likely before they happen.

Root cause classification — A multiclass model identifies the primary driver of each predicted stockout: restocking delay, system error, demand spike, or something else. The classification informs the response. Restocking delay? Escalate to logistics. Demand spike? Adjust the forecast.

Proactive alerting — Field teams receive alerts for stores at high stockout probability. They visit with purpose, not randomly. They fix the root cause, not just the symptom.

The beverage distributor went from discovering stockouts in-store to predicting them days in advance and understanding why they happen. Proactive detection at SKU × store level. Root cause classification across 3+ categories. Automated alerting to field teams.

Making It Work

Deployment matters. We ran the full pipeline on GCP—BigQuery for data, Vertex AI for models, and a performance dashboard for accuracy and model health. The 23-week timeline covered requirements, data integration, model development, and production deployment.

The key is treating OSA as a prediction problem, not a logistics problem. Logistics is the response. Prediction is the trigger. You need both, but prediction comes first.

Conclusion

On-shelf availability fails when teams treat it as a logistics problem. The real leverage is data: unified sell-out and inventory data, demand forecasting at SKU/store level, root cause classification, and proactive alerting. Our beverage distributor engagement proved it—23 weeks, GCP deployment, from reactive discovery to predictive prevention. OSA is a data problem. Solve that, and logistics becomes the execution layer, not the strategy.