The clerk who can only read labels.
Walk into a shop and ask for "a fridge that doesn't make noise". The clerk takes you to the quiet fridges. Now picture a clerk who can only read labels: they look for a fridge with the word "noise" printed on the box, find none, and tell you the shop doesn't have it.
That is how most online store search boxes work. They compare letters: if the customer types "refrigerator" and the catalog says "fridge", there are no results. If they type "microwave oven" and the catalog says "microwave", often nothing either.
Every store stands on three pillars: the storefront (photos, product page, banner), the cart (payment and shipping) and search. The first two get investment every year. The third stays as it came out of the box, and the whole store leans.
Six ways search says "we don't have it" when the store does.
All of them show up in any store's search log. Each one is a sale that was already decided and went to a competitor.
A different word
The customer uses their word, the catalog uses the supplier's.
One letter off
On a phone, in a hurry, a lot of searches arrive with typos.
A different spelling
Hyphens, numbers written out, abbreviated units.
How people talk
"for", "that doesn't make noise", "gift for my dad": the search demands every word in the product name.
Found, but in the wrong order
The customer asks for the appliance and gets the cover, the bracket and the remote first.
The dead end
"No products found" and nothing else. No suggestion, no category, no contact.
Two techniques: understand the question and search by meaning.
1. Understand the question
Before searching, the search takes the sentence apart: what is the product, what is a measurement, what is a city, what is just a way of talking. "for" and "my" go away; "12k btu" becomes "12,000 BTU"; "220v" becomes voltage.
For that it needs your store's vocabulary: that "refrigerator" means fridge, that "air fryer" is the same as "airfryer", that someone typing "split" wants an air conditioner. We build that dictionary from your catalog and from what your customers actually type.
Once the question is understood, search becomes a filter: only air conditioners, only 12,000 BTU, only 220 V. The appliance comes before the cover.
2. Search by meaning
No dictionary covers everything. For the rest, the search uses a map of topics: each product gets an address on that map, worked out from its name and description. Things that mean the same end up in the same neighborhood, even with no word in common.
When the customer asks, the question gets an address too, and the search returns the nearest neighbors. "Gift for a barbecue lover" lands next to the grill, the knives and the apron, although none of them has "gift" in its name.
And there is a cut-off: we measure, on your catalog, how similar two random products already look by chance. The search only returns what sits above that. Better to say "we don't have it, try these" than to show just anything.
Not theory: it's live, with numbers.
Both techniques run in production on our projects. One in local search directories with 760 thousand businesses; the other in a technology news portal with 95 thousand articles. The numbers below are before and after, measured.
Local directories · understand the question
News portal · search by meaning
The directories are ours; the portal is a client, hence no name. A store has fewer items than either, which makes search faster and cheaper, not the other way around.
The search box is the store's cheapest market research.
Every search is a customer saying, in their own words, what they want to buy. Almost no store reads it. In our directories, the first change was simply starting to record what people type: before that, nobody knew.
Every month you get the list of searches that went nowhere, split three ways: you sell it, but it's written differently (we teach the search); found, but in the wrong order (we fix it); and you don't sell it, which is buying information for your team.
Searched and didn't find · September
Five minutes, six searches.
Open your store on a phone and run these searches. Tick the ones that worked. You don't need to send anything to anyone.
Your search, by your test
It starts with a free test and continues monthly.
We run 50 customer searches on your store (typos, synonyms, sentences, measurements) and send what came back for each one, with a screenshot.
We build the dictionary from your catalog and real searches: synonyms, spellings, measurements, brands and what is an accessory of what.
We switch on question understanding and search by meaning, with automatic fallback to the old search if anything fails.
Searches with no results, new terms, what customers ask for that you don't sell. We tune the search; the rest comes explained.
What people always ask.
My platform already has search. Why touch it?
The search that ships with the platform compares letters. It doesn't know that a fridge is a refrigerator, that 12k means 12,000 BTU, or that the customer wants the appliance before its cover. That vocabulary belongs to your store, and someone has to teach it.
Do I need to switch platforms?
In most cases, no. The new search runs alongside the store you already have, when the platform lets you replace or extend its search. The free test tells you whether yours does, and how.
Isn't AI search expensive?
It doesn't have to be. For the news portal we serve, preparing 95 thousand texts cost about US$ 0.15, once; day to day it costs cents per year. A store catalog is usually smaller than that.
What if search by meaning gets it wrong?
It has a cut-off measured on your catalog: below it, it returns nothing rather than filling the page with noise. Brand names and model codes still go through letter search, and if the AI service goes down the store falls back to the old search on its own.
Does this help me show up on Google?
Not directly: site search serves people already in the store, and results pages shouldn't even go to Google. But what it teaches helps. Searches with no results show which words customers use, and those words become category names, product copy and new pages.
What about AI agents that shop for the customer?
They also ask the store questions, more and more through its search. A search that understands "split 12k 220v under 3,000" serves both the customer and the agent. See the agent-ready store.
Searching by what things mean, not by their letters: the question and the products get an address on a map of topics and the search returns the neighbors. The technical name is semantic or vector search. More
The dictionary that tells the search that "refrigerator" is a fridge, that "12k" is 12,000 BTU and that the cover is an accessory of the air conditioner.
A search that ends in "no products found". It is the most valuable list the store has: a decided customer, saying what they want, leaving without buying.
Where this might be wrong.
Bad catalog, bad search
Search by meaning reads the name and the description. A product with no description, or with the supplier's text copied onto a hundred items, lands at a vague address on the map. Sometimes the first action is fixing the copy.
Names and codes
For "Brastemp BRM44" or a part number, letter search is still the best. That's why both live together: the code goes straight through, the sentence goes by meaning.
Measured on directories and a portal
The numbers on this page come from local directories and a news portal, not from your store. The free test measures your search before any promise.
The examples are made up
The appliance store in the specimen and the zero-result report are illustrative. Your numbers come from your store's logs.
Start with the free search test.
Send us your store's address. Within 24 business hours you get what 50 customer searches find on it today, with a screenshot of each, in plain language.
- SaborCidade and SaúdeCidade directory search: a battery of 40 queries before and after, and response time measured in production, September 2026 (our own project).
- News portal search by meaning: 95 thousand articles; preparation cost, similarity cut-off and related-article coverage measured on the real archive, September 2026 (client project).