Journal · Sessions
Reformulation chains as a proxy for intent failure
A popular definition of search success inside apps is “the session contained a result click.” It is easy to instrument and easy to game. Users who type a nickname, then a legal name, then a partial account number — and click only on the third try — look successful. They have done your synonym work. They are tired.
Reformulation chains are consecutive queries in the same session that look like the same job: edits, expansions, script switches, or a move from brand to generic. In-app search analytics should flag them. They are one of the cheapest proxies for intent failure once your session identifier survives a backgrounded phone. That last condition is not optional; naive session IDs die in mobile traffic. The Lab’s first fortnight exists partly for this reason.
How to recognise a chain without a PhD
You do not need embedding drama to start. Same session, short time gap, substantial character overlap, or an obvious script switch (Thai to English brand, or the reverse). Tag a sample by hand. You will find families: spelling repair, synonym hunt, constraint add (size, colour, station), and abandonment disguised as a new query that is actually a category browse after search failed.
A digital bank mini-app in our Instrument Bench measured success as any result click. After they logged reformulation flags, they saw nickname → legal name → account fragment as a repeating shape. Click-through barely moved when they added nickname synonyms. Time-to-first-useful-result did. That is the point of the proxy: it changes which number you celebrate.
False engagement
Some products count query count per session as “engagement with search.” In a catalogue app that number often tracks confusion. Report chain rate beside query volume. If both rise after a release, you probably made retrieval worse or advertising more sloppy, not search more beloved.
The third query in a chain is rarely curiosity. It is usually labour.
What to do once you can see chains
Synonym tables for high-volume nickname families. Empty-state forensics if the first query in the chain was a zero. Ranking checks if the first query returned results that were plausible but wrong — presentation bias can cause reformulation too; see the click gap note. Do not auto-rewrite aggressively without a human sample. Silent rewriting that turns a dish name into a random SKU trains users to distrust the box.
Limits of the proxy
Exploratory browsing produces query sequences that are not failure. Comparative shopping (“red”, then “red size m”, then a second brand) can be healthy. That is why taxonomy sits upstream. A chain tagged as exploratory should not hit the same alarm as a nickname hunt. If your team cannot tell those apart, sit taxonomy week before you industrialise a reformulation dashboard.
For the full sequence we teach, start at in-app search analytics or programs.