Labelling Recommendation Impressions Without Noise

An impression is only useful if you know the surface, the position, the slate identity, and whether the item was actually visible. Partial viewport rows and prefetch caches routinely pollute training sets when those fields are approximate.

Prefer a small set of stable labels over creative event names for every experiment. Experiment IDs belong in a separate field; surface names should outlive the campaign that introduced them.

Document leakage risks: if a “viewed” event fires before paint, ranking will learn to favour items that load fastest, not items people wanted. That looks like model quality until latency improves elsewhere and the pattern collapses.

A quiet instrumentation review often unlocks more ranking progress than another architecture experiment.