AI-Powered Search Inside Your Online Store: Fixing Zero Results

The most honest report in your store is the failed search log

Someone types a query into your store search, gets nothing back, and leaves. They told you exactly what they wanted, in their own words, and the site said you do not sell it. Often you do.

Shoppers who use site search tend to have higher intent than those who browse, so these failures are expensive. Yet most store owners have never opened the list of searches that returned zero results. It is the cheapest piece of ecommerce research available, and it usually points straight at the fix.

Why does store search return no results?

Group a month of zero-result queries into these buckets. For most stores, vocabulary and attribute problems dominate, and those are partly fixable without new software.

Fixes that need no AI at all

• Add synonyms for the top mismatched terms, including regional ones such as trainers and sneakers, or hob and cooktop
• Index the attributes people search by: size, capacity, material, compatibility
• Turn on typo tolerance if your search engine supports it; many platforms ship with it off or weak
• Normalise part numbers and measurements in the index
• For products you do not stock, show a helpful page pointing to close alternatives rather than an empty grid

These changes often clear a large portion of the zero-result list within a week. If your platform's built-in search cannot do them, a dedicated search engine usually can; we compare the options in our search infrastructure comparison.

Where AI product search genuinely helps

The remaining failures are usually about meaning. Vector search converts product descriptions and queries into embeddings, numerical representations of meaning, so a query about keeping a dog cool can match a cooling mat even though no words overlap.

Pure vector search has its own weaknesses. It is poor at exact matches, so a search for a specific model number might return similar-looking products instead of the exact one. That is why stores should use hybrid search: keyword and vector retrieval combined, with a ranking step that weighs both.

1. Query understanding. Extract structured filters from the query, so oak table under 400 becomes material oak, category tables, price below 400.
2. Hybrid retrieval. Run keyword and vector search together and merge the results.
3. Re-ranking. Order results using relevance plus business signals such as stock, sales and margin, within limits.
4. Fallbacks. When confidence is low, show the closest category and suggest refinements rather than a blank page.

A language model can do query understanding well, but it adds latency and cost per search. Many stores use a small fast model or pre-computed rules for common patterns and reserve the larger model for long, conversational queries.

An illustrative example

A garden and pet supplies shop with 9,000 products and roughly 4,000 searches a week finds that about one search in eight returns nothing. Grouping those queries shows the largest buckets are synonyms and plural forms, then attribute searches such as sizes and breeds, then long natural-language questions.

Synonyms and attribute indexing fix the first two buckets. Hybrid search is then added for the long queries, with query logs reviewed weekly for the first two months. The measure of success is not the zero-result rate alone, which can be gamed by always returning something, but search exits and conversion from search sessions.

What to measure

• Zero-result rate, tracked weekly
• Search exit rate: sessions that leave straight from the results page
• Click-through from results, and the position clicked
• Conversion rate of search sessions compared with non-search sessions
• Refinement rate: how often people search again immediately, which suggests poor first results

Returning irrelevant results instead of none makes the zero-result rate look better and the experience worse. Watch click-through and exits alongside it.

When an AI search upgrade is not worth it

Small catalogues of a few hundred products with good navigation often get more from better categories and filters than from any search upgrade. Stores where few visitors use search may not recover the cost. And if the product data is thin, vector search has little meaning to work with; it cannot find a waterproof jacket if nothing says it is waterproof.

Our earlier guide to AI-powered website search covers content sites and knowledge bases. For product catalogues, SpiderHunts usually starts with the log analysis above and the quick fixes, then scopes AI integration only for the queries that remain. That order keeps the budget pointed at actual failures.

Frequently asked questions