Flesch Reading Ease Score, What It Measures and What It Misses in AI Writing
Flesch Reading Ease is a 1948 formula that scores text from two counts only: how many words are in an average sentence and how many syllables are in an average word.
Flesch Reading Ease is a 1948 formula that scores text from two counts only: how many words are in an average sentence and how many syllables are in an average word. Higher is easier, and 60 to 70 is the usual target for general writing. It is a quick check on long sentences and long words, but it cannot see meaning, structure, accuracy or the stock phrases that make writing read as AI, and Google has said it does not use reading level scores to rank pages. In our own tests of AI blog drafts, the score moved a lot with word choice and sentence length, and it also marked down a model whose drafts had almost no AI tells. That is why Riten shows the score but does not count it.
What is the Flesch Reading Ease score?
Rudolf Flesch published the formula in 1948, in a paper called "A new readability yardstick" in the Journal of Applied Psychology. It takes two averages from a piece of text and turns them into one number:
Reading Ease = 206.835 minus (1.015 x average words per sentence) minus (84.6 x average syllables per word)
Look at the weights. Each extra syllable per word costs far more than each extra word per sentence, so long words pull the score down fast. Microsoft's help page for Word gives the same formula and says "This test rates text on a 100-point scale." It adds: "For most standard files, you want the score to be between 60 and 70." In Word you can see the score under Editor, then Document stats.
The usual bands, as published tables give them, with labels that vary slightly from table to table:
| Score | Reads as |
|---|---|
| 90 to 100 | Very easy (about 5th grade) |
| 80 to 90 | Easy |
| 70 to 80 | Fairly easy |
| 60 to 70 | Plain English (8th and 9th grade) |
| 50 to 60 | Fairly difficult (10th to 12th grade) |
| 30 to 50 | Difficult (college) |
| 10 to 30 | Very difficult (college graduate) |
| Below 10 | Extremely difficult |
For a sense of scale, the same source puts Reader's Digest at about 65, Time at about 52 and the Harvard Law Review in the low 30s.
How is it different from the Flesch-Kincaid grade level?
The grade level uses the same two counts but reports a US school grade instead of a 0 to 100 score. It came from the US Navy in 1975. J. Peter Kincaid and colleagues recalculated the formulas from tests of 531 Navy enlisted personnel at four technical training schools, who read 18 passages from Navy training manuals.
The grade formula is (0.39 x average sentence length) plus (11.8 x average syllables per word) minus 15.59. Microsoft suggests a grade of about 7.0 to 8.0 for most documents. The two scores run in opposite directions: a higher Reading Ease is easier, a higher grade level is harder. Both rest on the same counts, so both share the same blind spots.
What does the score miss?
The formula counts words, sentences and syllables. Nothing else goes in. That leaves a long list of things it cannot see:
- Meaning. Shuffle the words in a sentence and the score stays the same.
- Structure. Headings, lists and a clear answer up top do not count.
- Accuracy. A wrong claim in short words scores as well as a right one.
- AI tells. Stock phrases and filler pass as long as they are short. Our guide to the signs of AI writing in a store blog post covers what readers actually notice.
- Useful precise words. A material name like polyurethane costs points, even when it is the most useful word in the sentence.
Short words in short sentences can still be vague. "This great product is a top pick for many people" scores well and tells a shopper nothing. A buying guide is a good example of where the precise word matters more than the number, which our post on how to write a buying guide shows.
The right score also depends on the topic. In a 2021 Ahrefs study of search results for 15,000 keywords, the average score was 69.7 for food, 60.2 for marketing and 49.6 for engineering. A page about bearings will never read like a page about cookies, and it should not try.
Does Flesch Reading Ease affect SEO?
Not directly. In a January 2018 office-hours hangout, Google's John Mueller said that, as far as he knew, Google does not have basic algorithms that count words to work out a reading level. He added: "But it is something that you should figure out for your audience."
The ranking data points the same way. The same Ahrefs study, across those 15,000 keywords, "found virtually zero correlation between rankings and FRE scores." So the score is not the thing that matters; the reader is. A page your shoppers find hard to read still loses them, whatever Google does with it. For Google's wider position on AI-written pages, see does AI content hurt SEO.
What did our AI writing tests show?
Between July 30 and August 5, 2026, we scored AI blog drafts on the same topics, with our own Flesch calculation. Other tools count syllables a little differently, so they will give somewhat different numbers. The rounds were small, 5 to 7 posts each, and the competitor comparison was only two posts, so read these as directions, not laws.
| Drafts | Posts | Flesch range | Average |
|---|---|---|---|
| GPT-4o, raw drafts | 5 | 37 to 47 | About 41 |
| GPT-4o with our prose steps | Rounds of 5 to 7 | 56 to 64 | About 61 |
| Two of a competitor's posts | 2 | About 51 and 55 | n/a |
| A newer model we tested, raw drafts | 7 | 49 to 59 | About 55 |
| The same newer model with our prose steps | 7 | 52 to 61 | About 58 |
During our test (Jul 30 to Aug 5, 2026), Riten wrote with GPT-4o. Its raw drafts landed in the "difficult" band. Two of a competitor's posts scored higher, and the gap was not sentence length. The competitor's sentences were actually longer than ours. Our raw drafts lost their points on word choice: abstract words with Latin roots, where a short plain word would do.
So our prose steps target words first. A fixed list swaps stock words for plain ones: "utilize" becomes "use", "numerous" becomes "many", "approximately" becomes "about", "sufficient" becomes "enough", "optimal" becomes "best". Then a rewrite pass works through each block of text, and its output is rejected if it drops an image, a link or a heading. With both steps, the same model moved from about 41 to about 61.
The surprise came from a newer model we tested on a test server. Its raw drafts had almost no AI tells, the stock phrases and filler that give machine writing away. Yet even after the same prose steps, it averaged about 58, below the setup we used at the time. The reason was sentence length: about 19 to 21 words on average, against about 15 to 16 for ours. The formula punishes long sentences whatever they say.
The lesson: the score rewarded what it can count, shorter sentences and shorter words. It could not reward what it cannot count, fewer AI tells and more useful detail. That is why we read the drafts as well as scoring them.
Why does Riten show the score but not count it?
Every Riten article gets an SEO score with a breakdown. The breakdown includes "Readability (Flesch)", marked as informational only, never affects the score. Across our posts, the Flesch score sat in a narrow band and did not tell good posts from weak ones, so it does not move the number.
The plain-word swaps and the rewrite pass described above have run on every new post since August 2, 2026. Posts come in the answer-first shape: a direct answer up top, question headings, short paragraphs and a short FAQ. Quality Review checks structure before you see the draft, and Riten picks relevant products from your synced Shopify catalog or a product CSV, with their photos, links and alt text placed in the post.
What actually helps readers and AI answers?
If the score is a weak guide, what is a better one? The strongest evidence on AI answers comes from the GEO paper (Aggarwal and colleagues, KDD 2024). The researchers rewrote source pages nine different ways and measured how much each page showed up in AI-generated answers across a 10,000-query benchmark.
- Sources, quotes and numbers helped most. Citing sources, adding quotations and adding statistics gave a 30 to 40 percent relative improvement on the paper's main visibility metric.
- Simpler, more fluent wording also helped. The paper reports a 15 to 30 percent boost from its fluency and easy-to-understand rewrites. The easy-to-understand rewrite alone was about 14 percent on the main metric.
- Keyword stuffing hurt. It did worse than leaving the page alone.
One caution: the paper never measured Flesch. Its "easy-to-understand" method was a model rewrite, not a score target. Simpler, more fluent wording helped in this lab study; that is not the same as saying a higher Flesch score gets you cited. Our guide to answer engine optimization covers this research and what it means for a store.
Put together, the practical list for a store blog post is short:
- Answer the question in the first two sentences.
- Use the specific detail only you have: sizes, materials, who a product fits.
- Back claims with a source, a quote or a number.
- Prefer the plain word, but keep the precise one when it carries the meaning.
- Vary sentence length, and cut sentences that run long without need.
- Read it as a shopper would before it goes live. Our pre-publish checklist walks through that read.
Want to see whether AI answers cite your store today? Run the free AI visibility check.
FAQs
What is a good Flesch Reading Ease score for a blog post?
Microsoft suggests 60 to 70 for most standard documents, which is plain English for general readers. The right score also depends on the topic: Ahrefs found averages of about 70 for food, 60 for marketing and 50 for engineering. Match your reader, not a number.
Does a higher Flesch score help SEO?
Not directly. Google's John Mueller said in 2018 that, as far as he knew, Google does not use basic reading level algorithms, and a 2021 Ahrefs study found virtually zero correlation between rankings and the score across 15,000 keywords. Readers still leave pages they find hard to read, so readability matters for your audience even when the score does not.
Can a high Flesch score make AI writing sound human?
No. The score counts only sentence length and syllables, so it cannot see the stock phrases that make writing read as AI. In our tests, a newer model whose drafts had almost no AI tells scored lower than the setup we used at the time, because its sentences were longer. Edit for AI tells and useful detail instead of chasing the number.
Write blog posts that feature your products
Riten picks relevant products from your catalog, places their photos and links, and writes posts made for Google and AI answers.