In-app search teams often report a healthy click-through rate from results to product pages, then look confused when add-to-cart does not move. The drop is not mysterious. It is usually one of four leaks, and mixing them into a single “search conversion” number guarantees the wrong fix.
1. Query class, not average CTR
Split queries into branded, category, attribute, and long-tail garbage. Branded queries (“brand x milk 2l”) should convert. Long-tail typos should not be averaged into the same KPI. If branded CTR is fine and long-tail is dumping people onto thin listings, you do not have a ranking emergency. You have an inventory or rewrite problem wearing a search hat.
2. Zero results and near-zero
Zero-result rate is obvious. Near-zero — two or three weak listings — is sneakier. Shoppers still click, then bounce. Marketplace App Analytics should count near-zero as a first-class event. In Thailand, local-language spelling variants make this worse; a catalogue that only indexes one transliteration will look “searchable” in English QA and fail in production.
3. Ranking versus inventory
A relevance model can faithfully rank the wrong shelf. If the top three results are out of stock, far from the buyer, or sold by a seller with brutal chat latency, clicks will happen and trust will not. Search Ranking Diagnostics, the eight-session programme, spends two sessions on this split alone: do not retune BM25 until you have checked stock truth and seller quality on the head of the list.
4. PDP bounce that search does not own
Sometimes search did its job. The PDP then fails: missing variants, a delivery promise that collapses at address entry, reviews that are clearly about a different SKU. If you only instrument search_click and never pdp_meaningful_view (scroll past fold, variant select, or chat open), you will punish ranking for a page that was never ready.
The practical move: draw a four-box tree, assign an owner to each box, and refuse a single “search conversion” OKR until the tree is populated. The Signal Map places this under Discover → Evaluate. The longer critique version lives in Search Ranking Diagnostics.