Fixing Zero-Result Search Complaints in E-Commerce

Of all the ways site search can fail, the zero-result page is the one that gets noticed. A customer types a query, the page comes back empty, and someone forwards a screenshot to the merchandising team. It is the most visible search failure, and because it is visible, it gets attention. That attention almost always goes to the wrong place.

The usual response is to open a ticket with the search vendor, ask why the query returned nothing, and tune a setting: add a synonym, loosen a match rule, switch on query relaxation. Sometimes that fixes the specific query in the screenshot. It rarely fixes the pattern, because the screenshot is one instance of a structural problem, and tuning one query at a time treats symptoms in the order they happen to be reported. The higher-leverage move is to read the zero-result log as a whole, identify the structural pattern underneath it, and fix the cause at the catalog or query-understanding layer.

The log you already have and probably do not read

Every major search platform records zero-result queries. Algolia reports a "No Results Rate" and a list of searches without results (Algolia Search Analytics docs). Coveo's dashboard shows the count of search events that returned nothing, plus a list framed explicitly as content gaps (Coveo docs). Constructor surfaces a "Top Zero Results" tab (Constructor docs). Bloomreach populates a "No Search Result Keywords" report (Bloomreach docs). The data is sitting in the dashboard regardless of which vendor a brand runs.

What happens to it is the problem. Most no-results pages are built as dead ends: 68 percent of e-commerce sites present a no-results page that offers nothing beyond generic search tips, no recovery path, no suggestions, no alternative (Baymard Institute). If the page itself is treated as an afterthought, the log behind it tends to be too. The query that produced the empty page gets logged, and then it sits there.

The rate is not trivial. Vendors set a best-in-class target below 2 percent (Algolia). The academic comparison is instructive: in general web search, queries that return nothing make up under 2 percent (Altingovde et al., 2012), but in a specialized corpus the rate climbs, reaching 15 percent in a search engine as large as PubMed (Dogan et al., 2009). An e-commerce catalog is a specialized corpus with a specialized vocabulary, and it behaves accordingly.

This matters because search traffic is the traffic that converts. In a study spanning 609 million shopper searches across 113 retailers, shoppers who used site search made up 24 percent of visitors but drove 44 percent of revenue, and converted at 2.5 times the rate of non-searchers (Constructor, 2025). A zero-result page is a high-intent visitor hitting a wall. The log is a record of demand the site failed to meet.

Four causes that look identical and are not

Every zero-result query looks the same in the dashboard: a string and a count. Underneath, they fall into four structural categories, and the fix for each is different.

The catalog gap. The customer wants something the store does not sell. This is the one case where the empty page is the correct answer, and it is also a signal: a recurring zero-result query for a product the brand does not carry is a merchandising and assortment input, not a search bug. The leading vendors name this explicitly as one of the two root causes of an empty result, the case where the answer is to consider expanding the catalog (Algolia).

The attribute gap. The attribute exists in the catalog (often as a filter facet) but free-text search cannot reach it. A search combining a refrigerator's capacity and style at a major home-improvement retailer returned zero products, even though the site offered both as filter facets and 45 products matched (Baymard Institute, 2026). The data was in the catalog. The search box could not see it. This category is common: 56 percent of sites rate as mediocre or worse on search experience, and 41 percent fail to fully support the basic query types shoppers actually use (Baymard Institute, 2026). The attribute gap is invisible to anyone looking only at the query string, because the product is right there in the catalog.

The synonym gap. The product exists and the attribute is indexed, but the customer used different words. This is the most-documented failure mode in e-commerce search. As far back as a 2014 benchmark, 70 percent of sites required shoppers to search using the exact terminology the site used, returning nothing for "blow dryer" when the catalog said "hair dryer," or for "all-in-one printer" when it said "multifunction printer" (Holst, 2014, Baymard Institute), and the current Baymard benchmark shows the problem persists across most sites (Baymard Institute, 2026). A vocabulary mismatch is not a search-quality problem in the algorithmic sense; the engine is working as designed. It is a controlled-vocabulary problem, and the fix lives in the synonym table, not the ranking model. The synonym and attribute gaps both live in the catalog rather than the algorithm, which is the broader case that catalog data quality, not the engine, is usually the bottleneck.

The query-understanding gap. The customer's intent was clear, but the query was malformed in a way the engine did not absorb: a typo, a singular where the catalog has plurals, or a compound word the engine did not split. Misspellings alone are a meaningful share of traffic, running at around a quarter of all e-commerce searches by vendor estimates (Clerk.io), and 69 percent of sites do not surface autocomplete spelling corrections for slightly misspelled terms (Baymard Institute). Compounding adds another layer in German, Dutch, and the Nordic languages, where a word like Hundehütte has to be split into Hunde and hütte to match a catalog that stores the parts separately, and even modern language models do not solve decompounding off the shelf (Algolia; Minixhofer et al., 2023). This category is the one most often "fixed" by toggling a vendor setting, and it is also the one where the toggle most often masks the others.

The distinction matters because the four causes demand four different owners and four different timelines. Treating all of them as "the search vendor's problem" sends every ticket to the team least able to fix three of them.

| Cause | Owner | Fix timeline | | --- | --- | --- | | Catalog gap (product not in inventory) | Merchandising and buying | Buying cycle | | Attribute gap (indexed as a facet, not as text) | Catalog and indexing pipeline | Weeks to a quarter | | Synonym gap (different vocabulary) | Synonym table (merchandising edit) | One to two days | | Query-understanding gap (typos, plurals, compounds) | Engine configuration | Days |

One symptom, four owners, four timelines. The dashboard shows none of this split.

What the dashboard will not tell you

Here is the gap that the tuning approach never closes. Every vendor reports the zero-result rate. None of the standard vendor reports decomposes that rate by cause.

The dashboards confirm this directly. Algolia frames empty results as having two possible reasons and leaves the classification to the reader (Algolia). Coveo lists the failing queries as content gaps (Coveo). Constructor's zero-results tab suggests adding synonyms (Constructor), which is a single hypothesis applied to a mixed population. Bloomreach offers query relaxation, a feature that finds related matches by loosening the match criteria to reduce the count of zero-result queries (Bloomreach). Relaxation is useful, but notice what it does: it suppresses the symptom by serving partial matches, which makes the rate go down without telling anyone why the queries were failing in the first place. The number improves. The structural problem is now hidden behind a partial-match curtain.

This is why two stores with an identical 10 percent zero-result rate can have nothing in common. A 10 percent rate that is mostly typos is a query-understanding problem solved with spell correction and autocomplete. A 10 percent rate that is mostly attribute gaps is a catalog-indexing problem that no amount of synonym tuning will touch. A 10 percent rate that is mostly catalog gaps is a buying decision wearing a search costume. The aggregate number is identical. The work is entirely different. A vendor report that gives the rate and stops there has diagnosed a fever without saying whether it is an infection or a sunburn.

The diagnostic, not the dashboard

Reading the log structurally is not exotic, and the procedure is short. Pull a sample of recent zero-result queries, say the last 30 days. Classify each one into the four buckets: catalog gap, attribute gap, synonym gap, query-understanding gap. Count the buckets. Fix the largest first. A few hundred queries, classified by a person who understands the catalog, is enough, because the same patterns recur long before the sample gets large.

None of this is improvised. It is the standard methodology of search log analysis, formalized long before e-commerce search existed: collect the interaction data, prepare it, and classify it into categories (Jansen, 2006). The same manual-classification approach sorts web queries into intent categories by drawing a random sample and coding each one by hand (Jansen, Booth & Spink, 2008), and small, well-defined category sets characterize e-commerce search behavior across millions of queries (Sondhi et al., 2018). The buckets are the structure the dashboard's aggregate hides.

What the structure reveals is often the opposite of the initial complaint. The zero-result page that triggered the ticket is frequently not the largest bucket. A merchant who looked at the data before discontinuing a product discovered it was the single most-searched term on the site; it was simply never in stock (Coveo). The screenshot in the inbox is one query. The log is the pattern, and the pattern points to a different fix than the one the screenshot suggests.

The log even under-reports the problem. A failed query is rarely the end of the session; well over half of e-commerce queries occur within a reformulation chain, where the shopper rephrases and tries again (Hirsch et al., 2020). Shoppers who reformulate and eventually succeed never generate a complaint, but they generate friction, and the zero-result that started the chain is the root of it. The aggregate rate counts the empty page. It does not count the abandoned session three reformulations later.

The zero-result page is the visible symptom, and the visibility is an advantage: it is the cheapest signal a search system produces. The mistake is treating it as a list of queries to fix one at a time. Read structurally, against a graded judgment set built on the brand's own queries and its own catalog, it is a map of exactly where search is leaking and which layer the leak lives in. Building that judgment set is the same first move as measuring whether search works at all, since an unmeasured system is one whose failures cannot be assigned to a layer.

The TensorOpt sample diagnostic, built on the public WANDS product-search dataset (Chen et al., 2022), includes a worked zero-result decomposition: the four causes, the method for classifying a query log into them, and what the breakdown looks like in practice. Download the sample report to see the decomposition before the next vendor ticket goes out.