Query Reformulation: A Search Relevance Signal

A shopper searches "leather jacket," gets a page full of results, and immediately searches again for "leather jacket women." Nothing broke. The first query technically returned products. But that second search is data — it says the first result set didn't work, even though no error, no empty page, and no zero-result rate would ever flag it.

That's query reformulation: a shopper searching again, shortly after, with different wording. It's one of the few relevance signals that catches failures a results page can look fine on.

What Query Reformulation Actually Looks Like

Reformulation is a same-session pattern — a shopper's second search closely follows their first, usually within seconds, on a clearly related query. A few concrete examples across catalogs:

Fashion: "jacket" → "denim jacket" (the shopper narrows because the first page was too broad).
Beauty: "foundation" → "base makeup" (a vocabulary mismatch between how the shopper and the catalog describe the same product).
Electronics: "iphone case" → "iphone 13 case" (the shopper adds the specificity the first query lacked).
Grocery: "organic gluten-free granola bars" → "granola bars" (the first query was too narrow for what the catalog actually stocks).
Marketplaces: "trainers" → "sneakers" (a straightforward synonym gap, more on that below).
Multilingual stores: the same query typed again in a second language the shopper also speaks, when the first attempt returned nothing usable.

Not every follow-up search counts. A shopper searching "wine" and then "wine glasses" a minute later is browsing, not correcting a failure — the second query is a new intent, not a retry of the first. Reformulation specifically means a *related* query standing in for one that didn't work.

Why Reformulation Matters More Than Zero-Result Rate Alone

Zero-result rate catches the failures that are easiest to see: searches that returned nothing at all. It's a real and useful metric, but it only catches one failure mode. A search-quality problem doesn't require an empty page — a technically non-empty result set that's the wrong products is invisible to zero-result tracking, and often invisible to click-through rate too, if the shopper doesn't click anything before trying again.

Query reformulation is tracked separately for exactly this reason: it's a behavioral signal that fires even when the first search "worked" by every system-level measure. A results page that returns three barely-relevant products needs the same attention as one that returns none — the shopper experience is close to identical, but only one of those cases shows up in a zero-result report.

Common Reformulation Patterns and What Each One Means

Reformulations aren't all the same failure. Grouping them by pattern points straight at the fix:

Narrowing (a modifier gets added — "jacket" → "denim jacket"). Usually means the first query's results weren't ranked well enough for the shopper to find the specific item among broader matches. The fix is ranking or filtering, not catalog data.
Broadening (a modifier gets dropped — "organic gluten-free granola bars" → "granola bars"). Usually means the catalog doesn't carry that exact combination of attributes, or the search treated the extra terms as required instead of preferred. The fix is often in how strictly filters are applied, not in adding inventory.
Synonym swap ("trainers" → "sneakers," "foundation" → "base makeup"). A vocabulary mismatch — the product exists, just described differently than the shopper searched. This is the pattern ecommerce search synonyms exist to close.
Spelling correction (the shopper retypes a corrected version of the same word). Points at a typo-tolerance gap rather than a relevance-ranking one.
Category pivot ("protein bar" → "protein shake"). Often ordinary browsing rather than a failure signal — worth excluding from reformulation counts so the metric isn't diluted by normal exploration.
Language switch (the same query retried in a second language on a multilingual storefront). Points at incomplete language coverage in the index, not a ranking issue.

How to Track Reformulation Without New Instrumentation

Reformulation is derivable from search logs a store likely already has: group queries by session, look for a second query within a short window (a minute or two is a reasonable starting point) that shares meaningful overlap with the first, and exclude the category-pivot cases described above. No new tracking code is required — the raw event is already "a search happened," logged twice in the same session.

The judgment call is in what counts as "related." Comparing overlapping words, checking whether both queries land in the same product category, or simply reviewing a sample by hand for a few weeks are all reasonable starting points. None require a formal NLP pipeline to be useful; even a rough first pass surfaces the highest-volume patterns.

Turning Reformulation Data Into Fixes

Once reformulations are grouped by pattern, each type points at a different, specific fix rather than a generic "improve relevance" backlog item:

High narrowing volume on a query → review default ranking for that query; the right product may already be in the results, just too far down.
High broadening volume → check whether facets or filters are being applied too strictly, and whether the catalog needs richer attribute tagging.
High synonym swap volume → add the specific pairs shoppers are actually using, checked against real query data rather than guessed in advance.
High spelling correction volume → test typo tolerance directly against the misspellings appearing in logs.
High language switch volume → check index coverage for that language rather than assuming multilingual support "should" already handle it.

This is the same discipline as measuring zero-result rate before deciding how big a search change is warranted — reformulation just catches the failures that a zero-result count misses because the first query didn't return an empty page, only the wrong one.

The Bottom Line

Query reformulation is a relevance signal that doesn't require a system error, an empty page, or even a missed click to fire — just a shopper who searched again because the first attempt didn't work. Grouping reformulations into narrowing, broadening, synonym, spelling, and language patterns turns a vague "search feels off" complaint into a specific, fixable list, the same way search abandonment data does for shoppers who give up entirely instead of retrying.

Curious what your own reformulation patterns look like? Book a demo and we'll walk through what your search logs already show.

Frequently asked questions

What is query reformulation in ecommerce search?

Query reformulation is when a shopper searches again, shortly after an initial query, using different wording — typically because the first search didn't return what they were looking for, even if it technically returned some results.

How is query reformulation different from a zero-result search?

A zero-result search returns nothing at all, which is easy to detect and count. Reformulation catches a subtler failure: the first search returned results, just not useful ones, so the shopper tries again with different words. Both are relevance failures, but only one shows up in a zero-result report.

Why do shoppers reformulate searches that technically returned results?

Because "returned results" and "returned the right results" aren't the same thing. A page full of loosely related or mis-ranked products often reads to the shopper as a failed search, even though the system didn't error or return an empty page.

Does every reformulation indicate a search failure?

No. A shopper moving from one related but distinct intent to another — "protein bar" to "protein shake" — is ordinary browsing, not a correction. Filtering out these category pivots keeps the reformulation metric focused on genuine retries of the same intent.

How can I measure query reformulation without new tracking infrastructure?

Group existing search logs by session, then look for a second, related query within a short window of the first. This works with search event data most stores already log — it doesn't require adding new instrumentation before the first analysis.

What's the fastest fix for a high reformulation rate?

It depends on the pattern. Synonym swaps are usually the quickest to fix once the specific word pairs are identified from logs; narrowing and broadening patterns typically need a closer look at ranking and filter logic rather than a single quick change.