Blog

Typo tolerance: what store search should forgive, and what it shouldn't

We compared 4.4 million words people typed on phones with the words they meant. Most typos are one key off, and over half land in a word's first four letters.

By Dagur Leo, founder of Findlane · 15 October 2026

Someone on a bus types runnign shoes with one thumb. If the store's search answers "No results", it has just told them the store doesn't sell running shoes. That costs sales: at eBay, shoppers who hit a search with no results bought at 0.63 times the rate of those who didn't (Singh, Parikh and Sundaresan, 2012).

Typo tolerance is the part of search that matches runnign to "running". Almost every search box claims to have it, but each draws the line somewhere different: how long a word must be, how many letters may be wrong, and which ones. Where that line sits decides which shoppers find the product. To see where it should sit, we looked at how people actually mistype on their phones.

How often searches have typos

Google said in 2021 that "every day, one out of 10 search queries is misspelled" (Google). Microsoft puts Bing at "about 15% of queries" (Microsoft Research), and Walmart's search team at "roughly 13% of user queries" (Walmart, 2024). None of them says how it counted, but they agree on the size: somewhere between one search in ten and one in seven.

Stores don't all cope. Among 60 of the largest US and European online stores, Baymard Institute found in late 2019 that "27% of sites won't yield useful results if users misspell just a single character in a product title" (Baymard). Its research notes that "misspelled queries are common, especially on mobile" (Baymard), and in 2021 it found 69% of sites didn't suggest corrections as people typed a slightly misspelled search, a problem it saw "more severe for mobile testers" (Baymard).

What 4.4 million words typed on phones show

In 2018, researchers at Aalto University asked visitors to a typing-test website to copy short sentences on their phones as quickly and accurately as they could. 37,370 people were in the study's analysis, which reported an average of 36.2 words a minute with 2.3% uncorrected errors (Palin et al. 2019). The researchers published the data. We lined up each typed sentence in the published data with the one it was meant to be, word by word: 564,090 sentences and 4.4 million words from 37,606 people. We counted only the mistakes people left in, which is what a search box receives once someone taps search. How we counted has the details.

About one word in 45 was left wrong, and long words more often. 2.2% of words had a mistake. For four-letter words it was 1.9%, for seven-letter words 3.8%, and for words of ten letters 6.1%. Product names and the words that describe them are rarely short. At these rates, a three-word search such as trail running shoes would carry at least one mistake about one time in eleven, not far from Google's figure.

Most mistakes are small. 76% of the misspelled words were one edit away from the right word: one letter wrong, missing, or extra, or two neighboring letters swapped. 92% were within two edits. That's close to what Fred Damerau reported in 1964 for misspelled terms in an IBM retrieval system, where "over 80 percent" were a single mistake (Damerau, 1964). A search that forgives one or two edits can find nearly every one of them.

Most of them are a neighboring key. Of the one-edit mistakes, more than half were the key next to the right one:

And more than half land in the first four letters. It's easy to picture typos at the end of a word, where a hurried thumb trails off. The data says otherwise. The first letter is usually right, but the second, third and fourth are wrong about as often as any later letter:

One mistake in nine is a space. People ran words together (doyou, superbowl) 8,586 times, and split them (slap shot, wide out) 4,176 times. On a store, that is tshirt for "t-shirt", or back pack for "backpack". Typo tolerance that works one word at a time doesn't see these at all.

Autocorrect hides some typos and makes others look real. Among native English speakers typing words of five letters or more, people whose phones autocorrected left about a third fewer mistakes: 3.1% of words against 4.9%. But 43% of the mistakes that got through were other real words, against 18% for people without autocorrect. Real-word mistakes in the data look like boots for boost, signing for singing, and free for fee. To a search box, boots looks like a perfectly good search.

Why "the first four letters" isn't enough

Shopify's built-in store search is a common example of a cautious rule. Its help center says results "include matches that differ from the search term by one letter or that have two letters in a different order", and that "the first four letters of a search term need to be entered correctly for typo tolerance to take effect" (Shopify Help Center). Merchants can't change it: "There's no setting that controls how broadly search matches."

We applied rules like these to the mistakes in our data, in words of five letters or more:

A rule that protects the first four letters leaves out three in five of the typos people actually make. Shopify's page describes one more step: when a search finds nothing at all, it "relaxes the search criteria", matching titles and product types with words one character off. That catches some of the rest, at a price its own example shows: "care bears" might find "card years".

Other platforms' built-in searches do less. WordPress search, which WooCommerce stores use unless they add a plugin, matches each word only as typed, though it may be part of a longer word (WordPress source). BigCommerce documents no typo tolerance, offers an optional "did you mean" suggestion, and suggests adding "commonly misspelled products" as product keywords (BigCommerce). Adobe's paid Live Search for Adobe Commerce tries "a maximum edit distance of 1", and only when a search finds nothing (Adobe).

The near misses typo tolerance doesn't catch

Typo tolerance compares one typed word with one word in your catalog. Plenty of misses aren't like that:

• Joined and split words. tshirt, t shirt and "T-Shirt"; backpack and back pack; pitbike and "pit bike". One merchant found pit bike had results and pitbike had none (Shopify Community). In a sample of misspelled searches at Flipkart, a large Indian online store, a space in the wrong place was part of more than a quarter of them (Kakkar et al., 2023). Languages that build compounds, such as German, Dutch and the Nordic languages, have far more of these.
• Words spelled by sound. In the same Flipkart sample, 37% of misspellings were spelled the way the word sounds, such as shart for shirt and metras for mattress. Some are one or two edits away; many are further, and need synonyms or a correction suggested from your catalog's words.
• Plurals and other word forms. boot and "boots" are one edit apart, knife and "knives" aren't, and in many languages the endings change more. A Dutch shop found konijn (rabbit) didn't find "Konijnen" (rabbits) (Shopify Community).
• Accents and special letters. Shoppers type creme, cafe and strasse for "crème", "café" and "Straße". A search should treat them as the same, without counting them as typos.
• Regional spellings. In the typing data, people wrote labor for "labour" and humor for "humour". On a store, it's color and "colour", gray and "grey".
• Units and symbols. 3 oz and "3 ounce", 200 gigabyte and "200 GB". In Baymard's late-2019 benchmark, 33% of the stores didn't support "even the most basic symbol or abbreviation searches for units" (Baymard).
• Codes written differently. A shopper typed bb600 for the SKU "bb-600" and got no results (Shopify Community).
• Real words that aren't the word meant. boots for boost, from the typing data. Only the results can show the search went wrong, which is why a no-results page should suggest a correction and a page of odd results should say what it searched for.

When typo tolerance finds the wrong thing

More tolerance isn't always better. Every edit a search forgives can also turn one product into another:

• Short words. boot is one letter from "boat", puma from "pump", case from "cast". That's why good typo tolerance allows no typos in words under four letters, one in short words, and two only in long ones, and why exact matches must rank above typo matches: someone who types boat should see boats first.
• Codes and numbers. iphone 13 must not find the iPhone 12. One merchant tested 20 product codes and found nine never returned the right product: 203333 returned "203343" (Shopify Community). Another searched for the SKU BBSS0297 and got BBSS0299 and BBSS0237 as well (Shopify Community). Numbers should match exactly, and SKU fields shouldn't take typos at all.
• Brand names. nike is one key from "bike", an example Flipkart's search team gives of a neighboring-key mistake. A brand one letter from a common word drags that word's products in, or the other way round, so brands should be able to opt out of typos.

The rule of thumb: forgive typos where a near miss is still the same product, refuse them where a near miss is a different one, and rank exact matches first so tolerance only helps.

How we counted

The data is the processed set from Aalto University's mobile typing study (Palin, Feit, Kim, Kristensson and Oulasvirta, MobileHCI 2019): people on phones who copied sentences from a typing-test website in 2018. We kept the study's validated test sections from phones with an error rate under 25%: 564,090 sentences from 37,606 people.

• Words. Each typed sentence and its target were lowercased, stripped of punctuation, and split into words, then lined up with a sequence alignment. Words typed and meant one for one were compared; two meant words typed as one (or the reverse) count as a space mistake. Words left out or added, unfinished words (a typed word that starts the meant one), and numbers aren't counted as misspellings.
• Edits are the optimal string alignment distance: a letter changed, added or removed, or two neighboring letters swapped, each one edit.
• Neighboring keys are keys within one and a half key widths on a phone's QWERTY layout.
• The first wrong letter is the first position where the typed word differs from the meant one.
• Typo rules were applied by the meant word's length. The table counts typos only: a search may still find some of the rest through other means, such as Shopify's relaxed search or semantic matching.
• Real words are those among the 50,000 most common English words in wordfreq. Autocorrect use is the study's own measure from each person's keystrokes. The study found autocorrect users left about as many wrong characters as others; counted by word, they left fewer mistakes, but more of them were whole other words.

The sentences were copied, not composed, and searches are typed with less care and nothing to copy from, so treat the rates as a floor. 32% of the participants weren't native English speakers; their error rates were within a few tenths of a percentage point of native speakers' with the same keyboard settings.

Test your own store's search

Ten searches, on your phone, take five minutes. Use your own product names:

1. A missing letter in a long word: hedphones, waterprof.
2. Two letters swapped: jakcet, hikign boots.
3. A neighboring key in the first four letters: jscket, bsckpack.
4. A wrong first letter: jiking boots.
5. A joined and a split word: tshirt, back pack.
6. A plural and a singular: boot and boots.
7. A word without its accent: creme, cafe.
8. A product code without its hyphen, and one digit off. The first should find the product, the second shouldn't.
9. A short word that's one letter from another product: does boat still show boats first?
10. Something you don't sell. Does the page suggest something useful, or just say "No results"?

Each search that finds nothing is a shopper who was told you don't have what they wanted. Your search analytics' list of searches without results shows which ones your shoppers actually hit.

How Findlane handles typos

Findlane's search follows the same rule of thumb:

• Typos anywhere in the word. One in words of four to six letters, two in longer words, with two swapped letters counting as one. A wrong first letter is tried only when nothing matches without it.
• Exact matches first. They rank above prefixes and typos, and numbers always match exactly.
• No typos where a near miss is another product. SKU fields and brand names can be set to take none.
• The other near misses. Accents never count as typos, and with an index's languages set, plurals match too. Two more settings help: compound words match words written together or apart, and did you mean says "Showing results for running shoes", or suggests a correction when nothing matches.

The typo tolerance settings explain each. The playground runs on a demo store, where rnuning shoes, hikign boots and jscket all find what they meant.