What are recommendations?

Recommendations are the app's findings turned into suggested actions: each one names a specific thing worth doing — refresh this creative, clear out this dead stock, win back these customers, investigate this CPM spike — grounded in the metric that triggered it. The open count is the queue of findings nobody has decided on yet. The engines only ever suggest; every status change is a human's.

Formula

open recommendations = COUNT(*) WHERE status = 'new'

Worked example

The feed shows 12 open items: three money-losing promotions, two low-margin SKUs flagged for pricing review, a dead-stock SKU suggested for liquidation, four marketing anomalies and two win-back segments. Each carries its supporting numbers and, where computable, an estimated impact in euros — the feed sorts by that, so the top of the list is literally the most valuable unread finding.

How Saldo Metrics computes it

Two engines write to canonical.recommendation. The core engine scans the semantic-layer views for commerce findings: promotions whose net effect is negative, SKUs selling at thin margin, dead stock worth liquidating, supplier cost increases, and dormant high-value customers. It does not generate reorder suggestions. The marketing engine adds the marketing set — creative refresh, budget reallocation, campaign scale and pause, geographic opportunities, list-health alerts, and the three overnight anomalies. Every row cites its supporting metrics, so a recommendation can be checked against the widget it came from. Rows deduplicate per finding — an open recommendation is not re-created daily — and the merchant moves status from new onward (acted on, dismissed); the KPI counts what remains new. Dismissing is not permanent: a dismissed finding that still holds on a later run can be raised again as a new recommendation.

Why it matters

Dashboards answer questions you ask; the recommendation feed asks the questions for you. Most of its findings are individually small — a €400/month promotion leak, one fatigued creative — but they are exactly the kind of thing nobody checks weekly, and they compound. An open count that only grows means findings are accumulating faster than decisions.

Common mistakes

  • Treating recommendations as instructions. They are metric-grounded hypotheses; the supporting numbers are attached so you can disagree with reasons.
  • Letting the queue age. Anomaly recommendations in particular decay fast — a CPM-spike finding from three weeks ago is history, not action.
  • Dismissing without reading the rationale. A dismissed finding that still holds can come back on a later run, so dismissing unread only postpones it; read the rationale and act on it or decide why it does not apply.

Where you see this in the app

The Recommendations page and its open-count KPI, and the anomaly subset in the Anomaly Monitor widget.