Surface and decision
We name the recommendation surface under review and the product decision the analytics must support—discovery, continuation, or catalogue launch.
Case notes
Short, anonymised notes from recommendation insight engagements—how the question was framed, what we measured, and what changed next.
How a note is structured
Every case note follows the same arc so product owners can compare engagements without inventing a new vocabulary each time.
Signals in play
We list the events already collected, mark which ones reach training or reporting, and call out leakage or missing viewport truth.
What the data could not say
Honest gaps sit beside findings. If a cohort was too thin or a label was approximate, the note says so before any ranking claim.
Next measurable step
Each note ends with one or two measurement changes the team can ship before the next model experiment—not a platform wishlist.
CTR rose while return rate fell
A home-shelf experiment celebrated a click bump that concentrated on items users already searched for. The scorecard shifted to save-and-return within seven days; the next slate change targeted discovery slots only.
New SKUs waited twelve days for non-promo slots
Popularity fallbacks kept empty shelves rare but delayed learning. The team added an early-signal clock for new items and a temporary exploration quota with clear measurement.
Prefetch events trained the wrong preference
Impressions fired before paint, so ranking favoured fast-loading rows. Viewport-true impressions and stable surface IDs restored interpretability for the next release.