How PlainSpell is made

Each section that shows a definition or a number ends with a source line: the source, its own date and a link. Word pages link to the Wiktionary entry they quote. The search figures name the engine, the country and the twelve months they average. The research pages give their method and limits and offer their data as CSV. If a line does not match its source, that is an error, and we want to hear about it.

Definitions, forms, pronunciations, homophones, variants, abbreviations and translations come from the English Wiktionary, read from the Wiktextract extract published by kaikki.org (the dump of September 2, 2026, extracted September 28, 2026). Wiktionary is written by volunteers and licensed CC BY-SA 4.0; PlainSpell keeps its wording, shows a selection of the senses (old, dialect and rare ones are left out on pair pages), shortens a long definition with an ellipsis (…), and credits it with a link on every page. The same licence covers what PlainSpell builds from it.

Spanish, French, German and Portuguese words are explained in English from the same English Wiktionary, and each page adds the definition from that language's own Wiktionary (dumps of September 1, 2026), shown in that language and marked as such.

How often a word is written comes from wordfreq 3.1, a frequency list that combines Wikipedia, film subtitles, news, books, web pages and social media. Its data is a 2021 snapshot and is not updated; PlainSpell shows it as times per million words. A word's rank counts the word forms written more often than it; wordfreq stores frequencies in steps, so words of equal frequency share a rank (receive shares its rank with 41 other words).

How often a spelling is searched comes from Microsoft Advertising (Bing) and Google Ads keyword data, read through DataForSEO on September 30, 2026. Each figure is the average month from September 2025 to August 2026 for the United States.

A word gets a page when Wiktionary gives it a meaning of its own. Every page is open to search engines and listed in the sitemap PlainSpell gives Bing, except words whose every sense is offensive or vulgar: they keep their page for a reader who looks them up, without ads and outside the sitemaps. A word is shown in the site's word lists and among the neighbouring entries when it also carries a fact PlainSpell measured (its written frequency, a measured misspelling or a searched pair) or readers found it in search, and at least two more facts (a pronunciation, a recording, its forms, an origin, an example, a translation; one more for a word in another language, whose own dictionary's definition counts). Inflected forms such as ran or cats, and obsolete or nonstandard spellings such as seperate, lead to the page of their word.

The sitemap PlainSpell gives Google lists an English word when its page carries at least five facts, one of them measured by PlainSpell, and a word in another language when it has an English meaning, its own dictionary's definition and a written frequency in its language. The other pages ask Google not to index them until they carry that much; Bing and other engines may. Proper names of places and people (Suriname, Snyder) have a page when they are in written use or readers found them in search, with their spelling, letters, syllables and pronunciation and Wiktionary's meaning; they stay out of the word lists and out of Google's sitemap. The months and the days of the week are ordinary words here, because their spelling is a common question.

A wrong spelling appears on PlainSpell only when people are shown to use it. Candidates come from three published lists (Wiktionary entries tagged as misspellings, Wikipedia's lists of common misspellings and the codespell dictionary), from the spellings Wiktionary records as obsolete or nonstandard forms of a word (counted only when a list above names them too, or after a check by hand) and from single-letter changes to the 30,000 most frequent English words that occur in written English. Each candidate and its correct word were measured on Bing, because Google reports most misspellings merged into the correct word's figure.

A candidate is kept when it is searched at least 10 times a month, or when a published list names it and it occurs in written text; when it is the less common form in search and at least ten times rarer in writing; when its share of searches is at least its share of published writing (a form written relatively more often than it is searched is a word or a name of its own); and when the correct word is an ordinary English word. Searches for brands, products, names and half-typed words were removed by rule and by hand after a review of the most searched candidates, and a wrong form that lists several possible targets is matched to the nearest one (two swapped letters count as one change). A pair of words where one is a measured misspelling of the other is shown on the word's page, not as a word pair. The share shown on a page is the wrong spelling's searches as a part of the searches for both forms.

A pair of words has a page when people search for it (“a vs b”, at least 10 times a month on Google or Bing) and the question is about the words themselves: they sound alike according to Wiktionary, they are two spellings of one word, one is written as one word and the other as two, or they look alike. Pairs that differ only in their inflection lead to the pair of their base words. Pairs that nobody searches remain as a line on each word's page.

Syllables are shown from Wiktionary's hyphenation when it has one; otherwise the count is taken from the pronunciation only when the marked syllables and the vowel sounds give the same number. Pronunciations are Wiktionary's, in the International Phonetic Alphabet, with the accent it names. The recordings are Wiktionary's links to Wikimedia Commons; pressing Listen fetches the file from Wikimedia. Written frequency is shown per million words and placed on a scale from rare to common. Rhymes are computed by PlainSpell: two words rhyme when their pronunciations in Wiktionary share every sound from the last stressed vowel to the end (American pronunciations are compared with American ones, British with British), and words spelled with the same letters in another order are listed beside them.

A rebuild keeps each word in its own language (an entry is filed under the language Wiktionary gives it, not under the edition it appears in), refuses records without a definition, and compares its counts of words, forms, misspellings and pairs with the previous build. The measured sets are kept with their raw responses, so a figure can be traced to the request that produced it. Every page is also checked by automated tests for missing values and for sources without a date.

PlainSpell is built with AI tools. They wrote the code that builds the database and the pages, the page designs, and the text of the research, list and help pages, which is checked against the data. They do not write the definitions, which are Wiktionary's, and they do not produce the figures, which come from queries against the database. Sentences about a single word are filled in from that word's record by fixed templates.

Bing is a minority of search in the United States, so the misspelling figures show the pattern rather than the whole volume, and volumes under 10 a month are not reported at all. Wiktionary is edited by volunteers and some rare entries are thin. The frequency list mixes American and British writing and stops in 2021, so it cannot show a recent word's rise or say which country uses a spelling. A regional label is Wiktionary's, not a style guide's ruling.

Report an error to hello@plainspell.com. A confirmed error is fixed at its source in the pipeline, the page changes at the next rebuild, and the correction is listed on the corrections page. The Wiktionary extracts are refreshed about every three months and the search figures once a year; each refresh changes the dates shown on the pages.