--- title: "Writing Sentences an AI Can Actually Quote: The 18 Word Ceiling" url: https://hostmy.blog/write-sentences-ai-can-quote/ date: 2026-09-15 modified: 2026-09-03 lang: en author: "Aditya Sharma" description: "Across 11,346 cited sentences, nothing longer than 18 words was quoted once. Here are five rewrites and the four edits that do all the work." categories: - "RankReady" image: https://hostmy.blog/wp-content/uploads/2026/09/hmb-card-1866-1024x538.jpg word_count: 1559 --- # Writing Sentences an AI Can Actually Quote: The 18 Word Ceiling Across 11,346 cited sentences a study could extract, the mean cited sentence ran 9.27 words. The 6 to 10 word band carried 45.2 percent of everything cited. Nothing longer than 18 words was cited a single time. Not rarely. Zero, out of a hundred and fifty three thousand. Open your best performing post and count the words in the first sentence. Ordinary blog writing runs well past 20 words per sentence. Every one of those sits above the ceiling. ## The ceiling is 18 words, and it behaves like a hard edge The distribution is worth reading carefully, because it changes what you do about it. | Sentence length | Share of citations | | --------------- | ------------------ | | 6 to 10 words | 45.2 percent | | Mean cited length | 9.27 words | | Over 18 words | Zero | A soft preference would show a long tail. This does not have one. Whatever mechanism selects quotable spans appears to stop at a boundary, and the boundary sits at 18 words. [The full distribution behind that ceiling](/18-word-ceiling/) is worth reading before you argue with it. That gives you a rule you can actually apply. Any sentence carrying a claim you want quoted must land under 18 words, and ideally between 6 and 10. ## Position matters as much as length A short sentence buried on line 400 is still buried. 41.9 percent of AI citations come from the first 30 percent of a page. Your best short claim belongs near the top, not after the anecdote and the context and the three caveats. Homepages, for the record, are cited around 4 percent of the time, which is [why your homepage is the wrong page to tune](/homepage-wrong-page-for-ai/). The work happens on article pages. So the rewrite job has two halves. Shorten the claim sentences, and move them up. ## Five rewrites, with the counts Here is the method applied to real sentence shapes. Word counts in brackets. **Before (29 words).** While there are many factors that influence how AI systems select sources, our analysis suggests that sentence length is one of the more significant and consistently observable variables involved. **After.** Sentence length predicts whether a passage gets quoted. (8) Across 11,346 cited sentences, the mean cited sentence ran 9.27 words. (11) Nothing over 18 words was cited once. (7) Three sentences, all under the ceiling, and the middle one carries the number and its subject together. **Before (30 words).** If you're running a WordPress site and you haven't yet checked whether your pages are actually reachable by AI crawlers, this is probably the first thing you should look at. **After.** Check crawler access before anything else. (6) A page returning 403 cannot be quoted. (7) Run one curl request per crawler and read the status code. (11) The conditional opener was doing no work. Cutting it moved the instruction to the front of the sentence. **Before (26 words).** Our plugin, which was built specifically for the problem of machine readability on WordPress, works alongside existing SEO plugins without creating conflicts or duplicating their output. **After.** RankReady runs alongside Yoast, Rank Math, AIOSEO and SEOPress. (9) It writes the AI-era layer only. (6) Your SEO plugin keeps the Article and breadcrumb nodes. (9) The subordinate clause in the middle of the original split the subject from its verb by fourteen words. **Before (29 words).** Sourdough starter typically takes somewhere between five and seven days to become active enough to use for baking, although this can vary depending on the temperature of your kitchen. **After.** Sourdough starter takes five to seven days to become active. (10) Warm kitchens are faster. (4) Below 18C, expect nine or ten days. (7) The qualifier did not disappear. It moved into its own sentence, where it is also quotable. **Before (15 words).** This means that it will not work in the way that most people currently expect. **After.** Content negotiation reaches coding agents only. (6) Under the ceiling and still useless, because `This` and `it` refer to something the reader cannot see. Length was never the only problem. Sentence length, from the citation record Sentence length, from the citation record Hard ceiling, nothing cited beyond 18 words Top of the most-cited band 10 words Mean cited sentence 9.27 words 153,425 citations, six AI platforms. The 6 to 10 word band carried 45.2% of everything. Three numbers set the target: where the mass sits, where the mean sits, and where the record simply stops. ## Four edits do almost all of the work Every rewrite above uses the same four moves. **Cut the conditional opener.** Sentences that start with `While`, `Although`, `If you`, `Given that` or `When it comes to` spend their first eight words before reaching the claim. Delete the opener, start at the claim. **Split at the conjunction.** Any `and`, `but` or `which` joining two complete thoughts is a sentence boundary you have not used yet. Two sentences of nine words beat one of nineteen. **Move the qualifier down.** Caveats are real and should stay. They belong in the next sentence, where they become independently quotable rather than dragging the claim over the ceiling. **Make each sentence resolve alone.** A sentence starting `This means`, `It does`, `They can` or `That is why` cannot be lifted, because the referent is in a sentence nobody took. Name the subject again. Repetition reads slightly worse to a human and enormously better to a machine. ## Do not write the entire page at nine words Worth saying, because the failure mode of this advice is obvious and unpleasant to read. A page of uniformly short sentences reads like an instruction manual. Nobody finishes it, and the people who leave were the audience. The working split on the pages here is this. Claim sentences, the ones carrying a fact, a number or a definition, go short and stand alone. That is also [how trust signals get read one line at a time](/eeat-seo-ai-answers/), rather than as a property of the whole page. Connective prose, transitions, examples and the parts carrying voice can run longer, because they were never going to be quoted anyway. Aim for the top of the page and the first sentence of each section. That is where both the length rule and the position finding point. ## Measure your own page in one command Guessing at sentence length does not work. Count it. `curl -s https://yoursite.com/your-post/ -o /tmp/p.html python3 - <<'PY' import re html = open('/tmp/p.html').read() body = re.sub(r'<(script|style)[^>]*>.*?', ' ', html, flags=re.S|re.I) text = re.sub(r'<[^>]+>', ' ', body) text = re.sub(r'\s+', ' ', text) sents = [s.strip() for s in re.split(r'(?<=[.!?])\s+', text) if len(s.split()) > 2] over = [s for s in sents if len(s.split()) > 18] band = [s for s in sents if 6 <= len(s.split()) <= 10] total = max(len(sents), 1) print(f'{len(sents)} sentences') print(f'{len(over)} over 18 words ({len(over)*100//total}%)') print(f'{len(band)} in the 6-10 band ({len(band)*100//total}%)') print('\nLongest offenders:') for s in sorted(over, key=lambda x: -len(x.split()))[:8]: print(len(s.split()), s[:100]) PY` Read the first percentage as your score. Above 40% means most of the page sits over the ceiling. Under 20% is a reasonable target after a rewrite pass. The offender list at the bottom is the actual worklist. Ten minutes per page, no tooling required. For [scoring a whole archive rather than one page](/measure-quotable-writing/), the same counting runs against a sitemap. ## Headings: sequential, and fewer than you think Structure carries its own effect, and it points in a direction most people find counterintuitive. Sequential heading order, H2 then H3 then H4 with no skipped levels, correlates with a 2.8x lift. That part is free. Fix any page that jumps from H2 straight to H4. Volume is the opposite. A page chopped into twenty sections is worse positioned than the same page with eight. That has a direct consequence for FAQ blocks, since every entry generates a heading. Cap them at four to six real questions rather than padding to twenty. More detail on that in [structured data in WordPress](/structured-data-wordpress/). ## What the plugin does and does not do here None of the rewriting above is automatable, and no plugin should claim otherwise. Sentence work is editing, and editing is you. The machine readable half is different. Summary blocks near the top of the page, FAQ markup, `speakable` selectors pointing at your cleanest sentences, Markdown copies and crawler rules are repetitive and easy to get wrong by hand. That set is what [RankReady](https://wordpress.org/plugins/rankready-ai-llm-seo/) handles. It is a WordPress AI SEO plugin, it runs alongside Rank Math, Yoast, AIOSEO or SEOPress rather than replacing them, and setup is about five minutes. What that changes is crawlability and how cleanly your page parses. Whether a model then quotes any given sentence is outside what a plugin controls, and I would not trust a vendor who told you differently. To check the result on your own pages, [the one platform that still shows you the line it took](/gemini-text-fragments/) is the place to look. ## Run the counter on your best page Copy the Python block, change the URL, run it. It takes thirty seconds and it gives you a number rather than an impression. Then fix the eight longest sentences on that page and run it again. That is the whole loop. Once the sentences are right, the question underneath them is whether machines can reach the page at all, which is layer one of [a full pass over all five layers](/five-layer-ai-seo-audit/). What percentage of your best page currently sits above 18 words, and would you have guessed that number correctly?