The shift most store owners are treating as an SEO problem
Google AI Mode is rolling out through Summer 2026. Most of the discussion is about discovery: will my products show up inside the AI answers? That matters, but there is a second effect that lands squarely in product and engineering territory, and it is easier to act on because you fully control it.
AI Mode answers a full question directly. A shopper types "a warm waterproof jacket for hiking under £150" and gets a relevant answer, not a page of links. Do that a few times and you stop compressing your intent into two nouns. You start typing the way you actually think.
That behavior does not stay on Google. It changes the input distribution hitting the search box on your own store.
The input your search box is about to receive
For twenty years, on-site search trained users to type like a database: "blue jacket," "face cream," "running shoes." They learned that a full sentence returned nothing, so they adapted to the tool.
AI Mode reverses that training. The queries start to look like this:
- "a moisturizer that isn't greasy for oily skin"
- "a lightweight day cream under £25"
- "krem do twarzy" (on a store listed entirely in English)
Each one carries intent, a constraint, sometimes a negation, sometimes a different language. None of them are keyword queries.
Why keyword search breaks on these
Default WooCommerce search, and most plugins marketed as "AI search," are keyword systems under the hood: token matching, fuzzy matching, synonyms, an inverted index, sometimes a chat wrapper on top. They rank by token overlap between the query and your product text. That model has predictable failure modes once the input is a sentence:
Negation is invisible. "A moisturizer that isn't greasy" contains the token "greasy." Token overlap pushes greasy products up, the exact opposite of the intent. The word "isn't" carries the meaning, and token matching does not model it.
A budget is just more tokens. "A lightweight day cream under £25" contains "under" and "25" as strings to match against product text, not as a numeric constraint on price. Keyword search has no notion that a budget was expressed.
Cross-language means zero overlap. "krem do twarzy" shares no tokens with English descriptions. Token matching returns nothing, even though you stock exactly what was asked for.
Synonym lists help a little, but they widen matching, they do not add understanding. They still cannot read a price out of a sentence or invert a negation.
This is not a chatbot
Worth stating for developers, because the two get conflated. A chat widget is a conversation surface: the shopper types, waits, reads a reply, and clicks through. On-site search is a discovery surface: the shopper types and expects to see products, immediately. The AI Mode habit is about querying, not chatting. The fix for it is better search, not a bot bolted onto the corner of the page.
What "AI-ready" on-site search actually does
The phrase "AI search" has been stretched to cover almost anything, so it helps to define it by behavior. An AI-ready search box needs to:
- Understand a full sentence, not just isolated tokens.
- Read constraints out of free text: a price like "under £25," attributes like "lightweight" or "for sensitive skin."
- Handle negation, so "isn't greasy" removes the wrong products instead of ranking them higher.
- Match across languages, so a query in one language finds products described in another.
Mechanically, that comes from a different category than keyword matching: comparing the meaning of the query against products using vector embeddings, plus an intent-parsing step that pulls structured constraints (like a price ceiling) out of the raw text before ranking. You do not need the internals to evaluate it. The behavior is testable from the outside.
A behavior matrix you can hold in your head
| Query | Keyword search | Semantic search |
|---|---|---|
| "a moisturizer that isn't greasy" | ranks greasy products up (token "greasy") | returns oil-control and matte formulas |
| "a lightweight day cream under £25" | matches "under"/"25" as words, ignores budget | reads the price, filters to it |
| "krem do twarzy" (EN store) | no token overlap, no results | matches meaning across languages |
The one-minute test
You do not have to trust a label. Run this on any store, whatever search it uses:
- Type a full sentence with a use case: "a warm jacket for hiking in the rain."
- Type a sentence with a price in plain text: "a lightweight moisturizer under 25."
- Type a negation: "a moisturizer that isn't greasy."
- If you sell into more than one language, type a query in a second language.
If two or more return nothing useful, the search is keyword-based. That is not a criticism of the store, it is the category the tool belongs to, and it is the category AI Mode is about to make feel broken.
For developers choosing or building search
Two takeaways. First, do not take "AI search" at face value; the label sits on both categories, so test with sentences, constraints, and negations. Second, on-site search is a conversion surface, not a nice-to-have. A shopper who searches is telling you exactly what they want, and a "no results" for a question they would say out loud is wasted intent that never even registers as a failed search. It looks like a normal exit.
Where I'm coming from
I build semantic search for WooCommerce and WordPress (Queryra), so I see this gap daily. Rather than argue it, there is a public demo you can run the test against without signing up, including the cross-language and price cases above:
The full write-up is the canonical source for this post:
queryra.com/blog/google-ai-mode-woocommerce-search
If you are evaluating on-site search for a store right now, run the one-minute test first. It will tell you more than any feature list.
Top comments (0)