AI Search Synonyms, Misspellings, and the Zero Results Page
Shoppers search the words they know, not your product names. How AI search handles synonyms and typos, where it still returns nothing, and how to test with a real query list.
Your search tool demos beautifully on your own vocabulary. Try the shoppers' vocabulary instead. Sweatshirt becomes hoodie and hoody. A waterproof jacket is a rain coat. Someone types trainers on a US store, soothie on a baby store, or the brand name they half remember with two letters wrong. Every one of those is a buying intent with money behind it, and a search box that answers no results to a synonym is a shop assistant saying we do not sell that while standing in front of it.
AI search genuinely improved this. Semantic matching catches a lot of synonym and typo traffic that keyword search dropped. It did not finish the job, and the failures moved: now the box returns something for everything, including confident, irrelevant results for queries it should have handled with a redirect, a category, or an honest broadening. The audit has to change from did it return results to did it return the right ones.
Build the query list from your own logs
Start with the zero results report and the top 500 site searches, not with a brainstorm. Mark each query: exact product language, synonym, misspelling, attribute language like waterproof or gift for dad, and problem language like shoes for standing all day. That last group is where semantic search earns its money and where your catalogue copy usually fails it, because product pages rarely describe the job the product does in the words a shopper uses for the job.
Test the list monthly and keep the failures verbatim. A failure list that says synonym issues is useless. A list that shows hoody returned two results while hoodie returned 140 is a work order: a synonym rule, a redirect, or catalogue copy that uses both words somewhere honest. Most stores find their top 30 failed queries cover a disproportionate share of lost search sessions, and fixing thirty queries is a week, not a programme.
Judge the something it returns
For each test query, grade the first screen: right products, mixed, or confidently wrong. Confidently wrong is the new failure mode to hunt. Semantic engines will return running shoes for dress shoes if the catalogue is thin, rather than admit a gap, and a shopper who sees that stops trusting the box and goes back to menus, or leaves. Where the catalogue genuinely lacks the product, a broadening message that says we do not carry X, here is the closest category is better retail than a fake match.
Check facets and sorting on search results while you are there. AI search often bypasses the merchandising rules that govern category pages, so a searched rail can lead with low margin, low stock items your category pages would never lead with. The search results page is a shop floor. It should be merchandised like one, with stock and margin sense applied after relevance, not ignored in its name.
Fix in the cheapest layer first
Synonym rules and redirects are cheap, instant, and reversible. Catalogue copy changes are slower but compound, because they also feed external search and AI answers. Model tuning conversations with the vendor come last, armed with your graded query list. Vendors respond to a spreadsheet of real failed queries with session counts attached very differently than to a feeling that search could be better.
Keep the query list as a standing asset. Re-run it after catalogue overhauls, seasonal range changes, and search vendor updates. Search quality drifts as the catalogue drifts, and the zero results page is where that drift collects. A store that watches it monthly spends small money constantly instead of large money on a replatform that inherits the same untested queries.
Related reading: Glossary, About, When AI Search Returns Nothing: Synonyms, Slang, and Zero Results.
Frequently asked questions
How is AI search different from keyword search on synonyms?
Keyword search needs an exact or stemmed match, or an explicit synonym rule. Semantic search represents meaning, so hoodie and sweatshirt can match without a rule. That helps a lot, but it can also over match, returning nearby but wrong products with confidence. Both failure types belong in your test results.
What should a zero results page do?
Never dead end. Offer the closest category, spelling or synonym suggestions, popular products, and a way to browse. Track how often it is shown and for which queries. A zero results page with no exits converts a fixable vocabulary problem into a lost session.
Where do I get a good test query list?
From your own search logs: the top searches by volume, the zero results report, and customer service chat transcripts where shoppers describe products in their own words. Refresh it seasonally. Thirty to a few hundred real queries beat any synthetic list.
Should I add synonyms as rules if the AI handles them?
For your highest value failures, yes. A rule is deterministic: it works every time, instantly, and survives vendor model updates. Use rules for the proven money queries, let the model handle the long tail, and keep grading both with the same list.