Readability checker that shows which sentences to fix

What this readability checker measures

Paste text, or give it a URL, and it scores how hard the text is to read. You get five standard formulas, the sentences that run too long, every likely passive verb, and a check on whether the page opens with an answer short enough to quote. The second tab is a keyword density checker. It shows which words and phrases the text repeats most, and where your target keyword appears.

Start with the limit. A readability formula counts. It counts words per sentence and syllables per word, or letters per word, and puts those counts into an equation fitted decades ago to reading tests. It does not read. It cannot see whether a sentence is true, whether the order makes sense, or whether your reader already knows the jargon. William DuBay’s review of the research puts it plainly: the formulas use only “surface features” of a text^[4]. The same review adds the other half. Those surface features have stayed the best simple predictors of how hard a text is to understand.

So treat the score as a pointer. It tells you where to look: the longest sentences and the densest words. Your judgement, or better, a real reader, decides what to change.

Pasted text is analysed in your browser and is never sent to us. A URL is fetched once by our server. We read the main content, skip navigation, footers, tables and code, and send the text back to your browser to score. Nothing is stored.

Sources: Brysbaert^[15]; Microsoft^[10]; GOV.UK^[11]; Digital.gov^[13]. Brysbaert’s figure is a meta-analysis of 190 studies; most adults fall between 175 and 300 words a minute.

The Flesch reading ease and Flesch-Kincaid formulas

Rudolf Flesch published the reading ease formula in 1948^[1]. In 1975, J. Peter Kincaid and colleagues re-fitted it as a grade level for the US Navy, using test results from 531 enlisted personnel^[2]. Both use the same two inputs. That is why the tool above works as a Flesch reading ease calculator and a Flesch-Kincaid calculator at once.

Look at the weights. In reading ease, each extra word per sentence costs about one point. Each extra syllable per word costs 84.6. So long words move the score far more than long sentences do. The grade level is meant as a US school grade: 8 means a typical 13 or 14 year old could follow the text.

Two examples from Wikipedia’s article on the tests, which our code reproduces exactly^[3]. “The cat sat on the mat.” has 6 words and 6 syllables in one sentence, so it scores 116. “The Australian platypus is seemingly a hybrid of a mammal and reptilian creature.” has 13 words and 24 syllables. It scores 37.5 for reading ease and 11.3 for grade level. The scale tops out at 121.22, for text made of one-syllable, one-word sentences. The grade level bottoms out at minus 3.40.

Three more grade formulas

The tool also runs three older formulas. Each estimates a US grade, and each weighs the text a little differently.

Sources: Gunning Fog^[5]; SMOG^[6][7]; Coleman-Liau^[8][9].

The four grade formulas rarely agree exactly. A spread of two grades is normal. A bigger gap means the text is unusual, for example very short sentences full of long words.

Reading ease score bands

Flesch grouped scores into bands, each matched to a US school level. This is the version most tools show, from Wikipedia’s table^[3]. The checker above marks the band your text falls in.

For a sense of scale, Wikipedia gives a few published examples: Reader’s Digest scores about 65, Time magazine about 52, and the Harvard Law Review in the low 30s^[3].

What reading level should web content aim for?

There is no single right answer, and anyone who gives you one number is simplifying. Here is what the main guidance says.

• Microsoft, for Word’s built-in statistics: “For most standard files, you want the score to be between 60 and 70.” For grade level, “the recommended score is around 7.0 to 8.0”^[10].
• GOV.UK makes plain English mandatory for every page, and asks writers to “split up sentences that are over 25 words long”. It notes that 1 in 6 adults in England have very poor literacy skills^[11].
• Digital.gov, which now hosts the US government’s plain language guidance, refuses to set a grade: “Only write for 8th graders if your audience is, in fact, an 8th-grade class”^[14].

Our reading of that: write as plainly as the subject allows, for the reader you actually have. A page for tax lawyers can use tax terms. A page anyone might need, such as a returns policy or a pricing page, should read at about grade 8 or lower. Use the reading level checker above to find the sentences that push the grade up. Then decide whether each one has earned its length.

Why readability helps AI answers quote you

Be careful with this one. No AI company publishes a readability score its systems prefer. We looked for one from OpenAI, Anthropic, Google and Perplexity, and found none. Google’s guidance on its AI features says the opposite of a magic number: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”^[16].

What we can say is mechanical. An answer engine builds its reply from passages. A passage that answers the question in one or two plain sentences can be lifted whole and still make sense. A passage that leans on the paragraph before it, or buries the answer in the fourth clause of a 40-word sentence, is harder to use. Nobody has published that as a ranking rule. It is simply how quoting works.

That is why the tool has a quotable answer check. It applies the same rule as the GEO Audit’s “Direct answer near the top” factor: the first paragraph after the main heading should be 40 to 320 characters. In URL mode it reads that paragraph with the audit’s own code, so the two always agree. It is our audit’s heuristic, not an AI vendor’s requirement.

Plain writing also helps the person who arrives from an AI citation and wants to check the claim. For the trust signals a model can actually parse, see E-E-A-T when the reader is a machine. For what is known about how assistants pick their sources, see how AI assistants decide which sources to cite.

Keyword density: why there is no ideal number

The second tab is a keyword density checker. It counts every one-, two- and three-word phrase and shows how often each appears. Add a target keyword and it reports the count, the density and where the keyword sits: the title, the H1, the first paragraph, the subheadings and the URL. Density here is uses divided by total words, times 100.

Now the caveat, which matters more than the number. Keyword density is not a ranking factor. In Google’s January 2023 office hours, John Mueller answered the question directly: “Google does not have a notion of optimal keyword density.” He added that “being explicit does matter”^[18]. Back in 2011, Matt Cutts said in a Google video that after the first mentions there are diminishing returns^[19][20].

What Google does define is the failure. Its spam policies list keyword stuffing, including “repeating the same words or phrases so often that it sounds unnatural”^[17]. So the tool only ever warns about over-use. It never tells you to add a keyword. The line it draws, about 3% for the target phrase, is our rule of thumb for text that starts to read as repetitive. It is not a Google threshold, because Google has none.

What is worth checking instead: is the topic named early, in plain words, where a reader and a crawler expect it? Does the page answer the questions around the topic, rather than repeat its name? Google’s own self-check asks whether you are “writing to a particular word count because you’ve heard or read that Google has a preferred word count”, and answers: “No, we don’t.”^[21] The same logic applies to keyword counts. Our guides to keyword stuffing and to LSI keywords and topic coverage go further.

Passive voice, and when to change it

The passive voice checker looks for a pattern: a form of “be” (is, was, were, been, being and the rest), up to two adverbs, then a past participle. It knows the regular -ed forms and more than 170 irregular ones, such as written, taken and built. So it finds “the page was updated”, “it is not often used” and “they’re paid weekly”.

It is a heuristic, and it misses some. It skips passives built with “get”, such as “he got fired”, and passives with no “be”, such as “the report, written in 2019”. It ignores “she’s finished”, because ’s there usually means has. Words like tired or interested count only when “by” follows, since “I was tired” describes a state. Read each highlight before you change it.

Change a passive when it hides who acts. Digital.gov is blunt about it: “Passive voice obscures who handles what and is one of the biggest problems with government writing”^[12]. GOV.UK asks for the active voice because it “puts the focus on the user and the action they need to take”^[11]. Keep a passive when the actor does not matter, as in “the site was built in 2019”, or is not known. Our guide to active and passive voice has examples of both.

How to make text easier to read

Work through this with the highlighted text above. Your ticks are saved in this browser only.

Check every page on your site

This tool checks one text at a time. The pages that need it most are usually the ones nobody thinks to paste: an old landing page, or a help article written in a hurry. The GEO Audit crawls your site and runs 31 checks on every page, including the same direct answer check this tool runs.

The free account audits 10 pages with an email address and no card. To rewrite a page, the draft editor on the Growth plan and above scores your draft against the live page as you write. Every limit is on the pricing page. For one page in depth, the AI SEO checker runs all 31 checks for free. If a URL here will not load because it redirects, the redirect checker shows every hop.

Questions people ask

Sources