The Query Intelligence Lab did not make our marketplace “smarter.” It made our Tuesday search review shorter, because we finally agreed what a session was.
What people actually said after the labs
Attribution styles vary on purpose. Some names are full, some are first name and city, some stay with industry only. One note includes a reservation. That is how the room felt.
I came for ranking features. I stayed because empty-state diagnostics showed our mini-app was failing on station names with tone marks stripped. Chiang Mai users were not “low intent.” They were misspelled by our normaliser.
★★★★★ · 9/10
Thai Query Normalisation Workshop. Dense. I wanted a take-home tokenizer; they refused, correctly. Left with a failure-mode list our NLP vendor could not ignore.
Client in grocery delivery. Our zero-result rate looked “healthy” until the Lab forced us to split navigational barcode scans from recipe browsing. The healthy number was an average of two unrelated jobs. Mild complaint: the instrumentation week assumed more SQL comfort than our PM cohort had. We caught up, but labs one and two were steep.
Signal Pair was the right shape. I am a PM; my colleague owns events. We argued through the entire ranking-bias week and still sit together for search reviews. The extra taxonomy clinic paid for itself when we retired three vanity charts.
★★★★☆ · 7.8/10
Instrument Bench for a banking mini-app. Facilitation was calm. I wanted more time on click models and less on event contracts; our engineers needed the opposite. Fair split, just not my personal appetite.
Case files
Two longer stories
Fashion marketplace · zero-result rate that lied
A Bangkok womenswear marketplace arrived with a zero-result rate of 4.1% and a mandate to “fix search.” Query Intelligence Lab week three asked them to tag empty-state queries by intent. Nearly half were colour+size strings for SKUs they had delisted but still advertised on Instagram.
Search was not failing to rank. Acquisition was sending demand into a catalogue hole. The capstone memo moved the work to merchandising and ad ops. Ranking experiments were paused for a quarter. They still send alumni notes when a campaign relaunches a dead SKU.
Fintech mini-app · reformulation as a distress signal
A Thai digital bank’s mini-app measured “search success” as any session with a result click. Reformulation chains were invisible. After the Lab’s session-success module, they found users typing a bill nickname, then the biller’s legal name, then a partial account number — three queries, one click on the last try, logged as success.
They added a reformulation flag and a nickname synonym table. Click-through barely moved. Time-to-first-useful-result did. The team’s own reservation: they still cannot distinguish curiosity taps from genuine bill pay on some result cards, so presentation bias remains in the report.