How error correction works

Accents, typos, half-typed words and dead ends: the four layers that stop a search from ending in zero results.

"jumepr", "runing shoes", "headhpones". In a store with the search that comes with the platform, this returns zero results and the visitor leaves convinced you don't have the product. Here there are four layers between the typo and the zero, and none of them needs configuring.

1. Accents and capitals don't count

The typed term and the catalogue text are normalised before they are compared: "CAFÉ", "cafe" and "café" are the same thing to the search. Nobody types accents on a phone in the middle of a purchase, and it shouldn't be necessary.

2. Typos inside the word

The engine tolerates swapped letters, one too many or one too few, and finds the word the customer meant to type. The tolerance grows with the length of the word, and that part matters:

• In short words, no tolerance at all. "cap" and "cup" are one letter apart and are not the same search; tolerating there would swap the visitor's intent instead of correcting it.
• In long words, one swapped letter is rarely another word, and the tolerance opens up.

It's this layer that makes "headhpones" return the headphones, without you writing anything.

3. While it's being typed, it's the same search

The search panel searches on every keystroke, but only treats the search as finished when the typing settles. And it recognises that "vibrat", "vibratros" and "vibrators" are the same person typing the same thing, even with a deletion in the middle.

This has two consequences you notice:

• The visitor never sees a "no results" halfway through typing because of a word that isn't finished yet.
• Your analytics count searches and not keystrokes. One typed search counts as one bToken and one term, not the twelve intermediate states it went through. Without this, the store's most searched terms would be half-words nobody typed on purpose.

4. The dead end, read by AI

When the three layers above aren't enough, and the search really would fall to zero, AI search on dead ends steps in (switch in Store → AI in the store, with a plan feature of its own).

What it does is reread what the person typed, work out what they wanted, and search for that. And it tells them so: the results appear with the note "Results for X, interpreted by AI from Y", with the original term in plain sight.

The note isn't a courtesy. A search that returns results different from what was typed, without saying so, looks broken; saying so, it looks attentive, and the visitor sees in a second whether that's what they wanted.

And on your side: typos don't pollute the reports

In the panel, a rare typo sits alongside the term it was trying to be, and you see both: what was typed and the term it refers to. It's what stops "vibratros" from showing up as demand of its own, and it's also what makes the store's suggestion list show only correctly spelled terms.

The criterion is deliberately conservative, and for good reason: burying a real term inside another one would hide genuine demand for something you don't sell. So it only groups what is rare, looks a lot like a far more common term, and sits at the distance of a typo rather than of a different word.

In Optimisation there's a card, "Typos corrected", with the variants the engine already groups on its own. It's there for you to know what does not need a synonym.

What correction doesn't do

Correction is for someone who misspelled the right word. It isn't for someone who correctly spelled a different word.

If customers search for "trainers" and the catalogue says "sneakers", there's no typo at all: they are two different words, and no distance in letters brings them closer. That is solved with synonyms, and the list of the ones worth having comes to you, suggested from the searches with no results.