
Only 12% of US shoppers say they find exactly what they want every time they search a store. That figure comes from a Google Cloud survey run by The Harris Poll in late 2022. The same study put the cost of failed searches at more than $234 billion a year in the US. Searchers are the buyers who already know what they want. Losing them hurts.
Ecommerce site search is the search box, results page and ranking logic that help shoppers find products inside a store. Online sales keep growing too. The Census Bureau estimated US e-commerce at $340.2 billion in Q2 2026, or 17.1% of all retail sales.
This guide covers what breaks, how to fix it and which AI features are worth the build.
A store’s search system takes a shopper’s typed or spoken query and returns matching products, categories or pages. It has three parts. Autocomplete suggests queries as the shopper types. The engine retrieves and ranks matches. The results page shows them with filters and sorting. Every part fails in its own way.
The engine is the part shoppers never see. Its ranking decides whether “red running shoes size 9” returns red running shoes in size 9. Or 400 unrelated sneakers.
Search affects conversions because searchers arrive with intent. In the US, 69% of shoppers name the search box as their main way to find products, ahead of browsing. A successful search led 92% to buy the item they wanted. After a failed one, 81% said they would probably leave and buy elsewhere.
| Finding from the 2022 survey of US shoppers | Share |
| Use the search box as their main way to find products | 69% |
| Bought the item after a successful search | 92% |
| Bought at least one extra item after a successful search | 78% |
| Had an unsuccessful search in November or December 2022 | 76% |
| More likely to buy elsewhere after a failed search | 81% |
| Avoid sites where search gave them trouble | 82% |
The survey covered nearly 13,500 adults in 14 countries. US shoppers reported about four failed searches a month. Every failed search is a buyer with a product in mind and no way to reach it.
Store search fails most on non-product queries and abbreviations. Compatibility and use-case searches come next. Baymard Institute’s 2026 benchmark of more than 170 sites and apps found that 56% have mediocre or worse search. Exact product names cause the fewest problems. Shoppers rarely search that way.
The query type study, updated on April 29, 2026, splits searches into eight types:
| Query type | Example | Sites with issues | Fix |
| Exact | “Sony WH-1000XM5” | 12% | Index model numbers and SKUs as their own field |
| Product type | “sofa” when the catalog says “couch” | 20% | Add synonym lists from real query logs |
| Feature | “waterproof jacket” | 39% | Map product attributes to searchable fields |
| Use case | “jacket for hiking” | 43% | Tag products by activity and occasion |
| Compatibility | “charger for iPhone 15” | 44% | Store compatibility data as structured attributes |
| Abbreviation and symbol | “65in tv” or “1/2 inch” | 54% | Normalize units and symbols at index time |
| Symptom | “dry skin cream” | 37% | Link problem terms to product benefits |
| Non-product | “return policy” | 66% | Index help pages and route those queries there |
Read your own logs before choosing a fix. Sort the last 90 days of queries by volume, then tag each with its type. The weakest type with the most traffic goes first.
The core fixes are useful autocomplete, typo tolerance, synonym support, a helpful zero-results page and filters that match the catalog. Autocomplete shows the gap well. Baymard found that 80% of sites offer it, yet only 19% run it well. Fix the basics before adding AI. They decide most failed searches.
Good results pages also need clear product cards. The guide to conversion-focused web design covers layout and calls to action.
Losing shoppers at the search box?
WebOsmotic audits your search logs by query type, finds the searches that end in zero results or exits, and fixes the worst ones first.
Three AI features add the most: semantic search, hybrid ranking and query understanding. A semantic engine matches meaning, so “something warm for a winter wedding” finds wool coats. Hybrid ranking blends that with keyword matching. Learning-to-rank models then reorder results using click and sales data.
| AI feature | What it does | When it pays off |
| Semantic (vector) search | Matches meaning through embeddings | Long, descriptive or use-case queries |
| Hybrid search | Merges keyword and vector scores | Catalogs with exact SKUs and fuzzy queries together |
| Query understanding | Turns “black dress under $100” into filters | Stores with rich attribute data |
| Learning to rank | Reorders results from click and sales data | High-traffic stores with steady query volume |
| Personalization | Adjusts ranking per shopper | Repeat buyers and large catalogs |
| Visual search | Finds products from a photo | Fashion and furniture |
Open-source engines now ship hybrid search as a standard option. OpenSearch hybrid search combines keyword and vector scores in one query. Teams that keep data in Postgres can store embeddings with pgvector. The comparison of Pinecone vs pgvector covers that choice. For ranking by shopper, see the guide to AI ecommerce personalization.
Pure vector search can miss exact SKUs and model numbers. Keep a keyword path for those, and test both on real queries.
Choose ecommerce site search solutions by catalog size, query volume and the control you need. Built-in platform search suits small catalogs. A hosted search service suits mid-size stores that want fast setup. A custom engine on OpenSearch or Elasticsearch suits large catalogs, complex attributes or strict data rules.
| Option | Best for | Setup time | Trade-off |
| Built-in platform search | Under a few thousand products | Days | Limited control over ranking and synonyms |
| Hosted search service | Mid-size catalogs, small teams | 2 to 6 weeks | Per-query or per-record fees grow with traffic |
| Custom engine with hybrid AI | Large or complex catalogs | 3 to 6 months | Needs an in-house or partner team to run it |
Headless stores call search through an API, so they can swap engines later without rebuilding the storefront or the checkout, as the guide to headless commerce development explains.
Measure five numbers every week, split by device. Start with search usage rate and zero-results rate. Exit rate from the results page tells you how often shoppers give up and leave, which is the clearest sign that ranking or the catalog data has failed them. Search conversion rate and revenue per search put a dollar figure on the box.
Set a target for the zero-results rate and review it monthly. A falling rate is usually the first sign the fixes are working.
Ready for search that understands your shoppers?
Our engineers build hybrid keyword and AI search on your catalog, tune ranking from your own sales data and measure the lift in conversion.
Lower is better, but no published standard sets one number for every store. Take a baseline from your own logs, since the right target depends on catalog size and query mix. Track it weekly, then review the 100 most common zero-result queries. Usually the fix is a synonym or a redirect.
No. Semantic search handles descriptive queries such as “gift for a coffee lover”, while keyword search stays more reliable for exact SKUs and model numbers. Hybrid engines run both and merge the scores. Exact searches fail on 12% of sites in Baymard’s tests, so a keyword path still earns its place.
Synonyms, typo tolerance and a better zero-results page take 2 to 4 weeks on most platforms. A move to a hosted search service takes about 2 to 6 weeks. A custom hybrid engine with learning to rank usually needs 3 to 6 months, including testing on live traffic.
It needs clean product titles, structured attributes such as size and compatibility, plus stock status. Query logs and click data come next. Compatibility searches trip up 44% of sites in Baymard’s benchmark. The cause is usually missing catalog data.
Yes, once the catalog passes a few hundred products. In Google Cloud’s survey, 69% of US shoppers named search as their main way to find products. Start with the platform’s built-in search, add synonyms from your logs and upgrade only when zero-result queries keep climbing.