Everyone selling AI search says "100+ languages". I run one, so I went and checked what that actually means on real shops. No demo data. Real catalogues, real queries, and I am publishing the failures too, because the failures are the part you need before you buy anything.
Five catalogues, five languages. Two worked. Three broke, and they broke in three different ways.
1. Polish query, English catalogue. Worked.
A Brooklyn furniture store, catalogue entirely in English. I typed "małe sofy" (Polish for small sofas). 21 relevant results, sofas, correctly sized.
This is the case that actually justifies the marketing. Your shopper does not have to speak your shop's language. That part is real.
2. Portuguese, on a store selling in Portuguese. Worked.
A store selling in Portuguese, catalogue in Portuguese. Brand queries landed cleanly: a brand name returned twenty results, all of them that brand. A query for a product variant put the right five on top.
This one worked, and it is worth saying that same-language search is not hopeless. It is just where the hard cases live, and the next three are all hard cases.
3. Spanish, and this one is a single letter.
"camisa" (shirt): 20 results, all of them shirts, all exact matches.
"camisas" (shirts, plural): the top result was a three-pack of undershirts.
Same shop, same second, one letter of difference. The products were named "Camisa manga...", singular. Query in plural, product in singular, and the match quality fell off a cliff.
The pattern held across the catalogue. "sandalias" and "franelas" worked, because those products happen to be named in the plural too. "toallas" (towels) did not: the only exact match was a set of clothespins for towels. The actual towels came back as approximate.
This one is on my list to fix. I am including it because you will not find it by testing "shirt" in English.
4. German, and this is the hardest case there is.
A wool shop, German catalogue. I typed "etwas Warmes für kalte Hände" (something warm for cold hands). Zero results.
The shop sells merino hand warmers. They were sitting right there.
German compounds the words. "Walkloden" is a specific fulled wool fabric, and it is one word. Query a language that builds words like that, against a catalogue written in that language, and you are in the hardest corner of this whole problem. Anyone claiming 100+ languages should be asked about this specific case.
5. Mongolian, Cyrillic script.
A business news site, about 4,000 records. I typed "цай", which is Mongolian for tea. 20 results came back. The actual article about tea culture was eighth. Second place went to an article containing "цайз", which means fortress.
Then I typed a surname, "Ганбаатар". That worked, first result, exact match. Then I added the initial, "Д.Ганбаатар", the way a Mongolian reader would actually write it. The right result fell to fifth. Two of the four results above it were other people whose names start with Д. The other two were not people at all, just articles that happen to contain words beginning with the same letter.
One initial and a full stop were enough to push the correct answer down four places and fill the space above it with noise. Script and naming convention both matter, and neither shows up in a demo.
What I would take from this
Twenty results is not a good sign. On three of these five, a full page came back and the right answer was buried or absent. Zero results is honest. A full page of near-misses looks like it worked.
One more thing worth checking, and it has nothing to do with language: on one of these shops a query returned four blog posts before it returned a single product. The shop had written articles about the thing, and the articles outranked the thing itself. If your store has a blog, ask whoever sells you search what stops that from happening.
So if you are evaluating search for a non-English shop, do this: pick ten queries your customers actually type, in your language, including plurals and compounds, and check whether the right product is first. Not whether results appeared. Whether the right one is first.
That test takes twenty minutes and it is the only one that matters.
Ready to fix your WooCommerce search?
Free forever, no credit card, no API key
Frequently Asked Questions
Does AI search work on non-English stores?
Partly, and the part that works is not the part most people expect. Searching in one language against a catalogue written in another worked well: a Polish query on an English furniture catalogue returned 21 relevant results. Searching in the same language as the catalogue is harder, because it runs into the grammar of that language. A Spanish plural against singular product names dropped the right answer off the top, and a German query phrased as a compound returned nothing at all on a shop that stocked the item.
Which languages broke in this test?
Spanish, German and Mongolian, for three different reasons. Spanish broke on plurals when the product names are singular. German broke on compound words, where the shop writes one word and the shopper types a phrase. Mongolian broke on naming convention: a bare surname matched first, but adding the customary initial pushed the correct article down to fifth place.
How should I test search on my own catalogue?
Pick ten queries your customers actually type, in your own language, and include at least one plural and one compound word. Then check whether the right product is the first result, not whether any results appeared. A full page of near-misses looks like success and is not. The test takes about twenty minutes.