How to Reduce Zero-Result Searches on an Ecommerce Site | Particular Audience
Five steps to find, fix and prevent zero-result and near-zero-result searches, from query logs to meaning-based search.
How to reduce zero-result searches on your ecommerce site
To reduce zero-result searches, export your search log, group the failing queries by cause, fix catalogue data gaps, and replace one-off synonym rules with meaning-based search that understands phrasing. Then show the closest relevant products instead of an empty page, and monitor the zero-result rate weekly so new gaps are caught early.
1. Find them
Pull the queries that return nothing or very little, ranked by how often they are searched.
2. Group by cause
Typical causes: misspellings, synonyms, descriptive phrases, products you do not stock, and missing attributes in the catalogue.
3. Fix the data
Add missing attributes and categories. Search can only find what the catalogue describes.
4. Search on meaning
Vector search answers many phrasing variations at once instead of one synonym rule per query.
5. Never show an empty page
Show the closest matches and related categories, and keep monitoring.
Measure the event the shopper actually completed
Define your denominator before reviewing a zero-result rate. A completed search, an autocomplete request and a filter refinement are different events. Decide whether you report by search attempt, session or unique term and keep that definition stable.
PA's ATS guidance warns that recording partial queries as SearchTerm events can inflate the apparent number of zero-result searches. Check instrumentation before attributing every empty response to the engine. Search-event documentation.
Ask whether a relevant, eligible product exists
Find an item that should satisfy the query in the current assortment and fulfilment context. If no such item exists, the problem may be an assortment gap rather than retrieval. Record this separately so a search team is not rewarded for inventing a match.
If a suitable product exists, follow its path. Is it in the source feed? Was it accepted and indexed? Are the needed attributes present? Does a store, stock or eligibility condition exclude it? Does the query express an identifier, a brand or a constraint the retrieval path mishandles?
Choose the repair that matches the cause
Correct a broken feed at the source. Repair identifier mapping where an exact product is missed. Improve descriptions or structured attributes when the item cannot be found by a legitimate property. Investigate spelling and synonym handling when different language expresses the same intent. Treat unwanted filter combinations as a navigation problem with a clear recovery option.
For a synthetic query such as “blue 2-litre jug”, returning a red cup may eliminate the empty page while ignoring the request. A useful recovery preserves the required product type and identifies which constraint cannot be met. If only another colour is available, explain that alternative rather than presenting it as an exact result.
Distinguish recovery from silent broadening
Keep exact results, suggested corrections and alternatives recognisable. The interface should explain when it has changed the query or relaxed a preference. Never silently relax an eligibility or safety-related condition. A truthful no-match response can be more useful than a confident but invalid answer.
Report success with relevance and customer outcomes
Track how many failed attempts are diagnosed, how many have relevant recoveries and what shoppers do next. Review grades for recovered products alongside conversion, abandonment and repeat reformulation. A falling empty-result rate accompanied by worse relevance is not a successful repair.
Use the diagnostic worksheet to classify a small representative sample first. Fix recurring causes, replay the same queries and then monitor a fresh sample to check whether the improvement extends beyond the queries used to tune it.
Illustrative example
Synthetic diagnostic: 100 completed search attempts, 12 empty responses. Empty-result rate=12/100=12%. Of 12 failures, 5 are catalogue gaps, 4 indexing failures and 3 retrieval failures. Potentially recoverable=7; this is a diagnosis count, not achieved lift. Fixing all 7 would leave 5/100=5% only under unchanged traffic and assortment, so do not present that as a forecast.
Downloads
- zero-result-diagnostic.csv (illustrative template)
Sources
- Getting Started With ATS, checked 2026-10-25
How Particular Audience does it
In production
- Top-10 North American Grocery Retailer: Fixing grocery site search (+65% More relevant catalog surfaced (Recall@120 0.485 → 0.800))
Frequently asked questions
What is a good zero-result rate?
It varies by catalogue and traffic. Track your own trend and the revenue of the queries affected rather than chasing a universal benchmark.
Is zero zero-results a sensible target?
No. Some queries request products the retailer does not stock or cannot currently sell in that context. Separate genuine catalogue gaps from avoidable retrieval failures. Success means more useful, truthful outcomes rather than a product list for every possible string.
Should we replace an empty page with bestsellers?
Bestsellers can be a clearly labelled secondary discovery option, but they are not evidence that the query was answered. Keep the no-match explanation and preserve essential constraints. Measure whether the recovery helps rather than counting any populated page as success.
How do we avoid misleading improvements?
Keep event definitions and eligible assortment consistent, then grade recovered results. Report zero-result rate with relevance, reformulation and commercial outcomes. Do not compare a completed-search denominator in one period with autocomplete requests in another.
Related guides
- Vector search vs keyword search: what's the difference for ecommerce?: Keyword search matches words; vector search matches meaning. Why that matters for long, descriptive and conversational shopper queries.
- How to evaluate your ecommerce site search: Test site search relevance with a fixed query set, graded results and head-to-head comparisons, including LLM-as-a-judge grading.
- Product-data requirements for AI ecommerce search: Prepare product data for AI search with a field checklist covering identity, variants, attributes, prices, stock, update behaviour and measurement joins.