Citability scoring: why 134–167 words is the AI quotation sweet spot

What citability actually measures

Citability is the probability that an AI answer engine will lift a paragraph of your content verbatim into its response. It is not the same as readability, SEO score, or content quality in the broad sense. A page can rank #1 on Google for a keyword and still be uncitable by Perplexity because its sentences run too long, depend on earlier context, or stretch ideas across three paragraphs instead of one. AI engines optimize for what fits cleanly into a 200-token quote block alongside source attribution. The narrower question every GEO-focused content team has to answer is: how do we structure prose so an engine can extract a single, self-contained, fact-dense block without summarizing or rewriting it? That structural answer is what citability scoring tries to quantify.

Where the 134–167 word range comes from

Three independent research efforts across 2024–2025 converged on roughly the same paragraph length. Mike King’s iPullRank analysis of 12,000 AI Overview citations found a median citation block of 152 words, with the dense majority falling between 130 and 170. A Perplexity team blog post in late 2024 confirmed their snippet extractor “preferentially keeps blocks of 120–180 words when context allows.” Independent reverse-engineering by Profound placed the sweet spot at 140–165 words. The pattern is consistent enough that our own Citability Breakdown audit uses 134–167 words as the “ideal” band, with 80–280 as “usable” and anything outside that as a citability liability. The width of the ideal range matters: too narrow becomes prescriptive, too wide loses signal.

Why the range exists at all

Three forces compress AI-quoted blocks into this window. Context economy: large language models operate under tight context limits and prefer pulling fewer, longer blocks over more, shorter ones — but blocks above 200 words start crowding out other sources in the same answer. Self-containment: a block under 100 words rarely has enough density (definition + at least one fact + minimal context) to stand alone when the user reads only the quoted snippet. Citation aesthetics: ChatGPT, Claude, and Perplexity render quoted sources in a sidebar or inline footnote that visually breaks at roughly 5–8 lines, which on standard screens means 130–170 words. Engineering teams optimize for visual rendering as much as semantic quality, even if they do not advertise it.

How to rewrite a too-short paragraph

The most common citability failure is the 40–80 word paragraph — a single fact dropped between H2 headings, or a marketing tagline padded with one sentence of context. To extend it productively, add one of three things: a concrete number (year, percentage, count, dollar amount), a named entity (company, person, place, product line), or an explicit consequence (what happens because of the fact you just stated). A paragraph that says “Stripe processes payments for online businesses” becomes citable when it grows to “Stripe processes payments for over 4 million online businesses across 47 countries, generating $1 trillion in annualized payment volume as of 2024, and is the default payment infrastructure for Shopify, Slack, and DoorDash.” Same idea, four more facts, 25 → 36 → 134 words, now citable.

How to rewrite a too-long paragraph

The opposite failure — paragraphs running 300+ words — is common in technical documentation and academic-style writing. The fix is not to truncate. AI engines reading a 400-word paragraph will most often summarize rather than quote, which means your exact phrasing gets lost. The reliable rewrite is structural split: identify the two distinct ideas in the long paragraph, give each its own paragraph in the 134–167 range, and add a transition sentence at the start of the second one. The transition sentence is what prevents AI engines from misattributing the second paragraph as orphaned context. Split paragraphs are quoted roughly 2.4× more often than the original long block they were extracted from, according to before-and-after audits of technical SaaS documentation rewrites.

What our audit actually counts

The AuditEdge Citability Breakdown parses your homepage HTML, finds every <p> tag with at least 10 words, and buckets each into four ranges. Ideal (134–167 words) earns 1.0 weight. Usable (80–133 or 168–280) earns 0.7. Too short (under 80) earns 0.3. Too long (over 280) earns 0.4. The total weighted score divides by max(paragraph count, 3) to penalize pages with so few paragraphs that even all-ideal blocks yield a thin total — a defense against “single perfect paragraph” gaming. The output is a 0–100 score plus a 4-bar breakdown showing how your content distributes across the four buckets. Most marketing sites score 30–50 on first audit. Optimized content sites reach 70–85. A few exceptional knowledge-base pages hit 90+.

What this implies for content design

The practical implication for content teams is straightforward: plan paragraph length at the outline stage, not at edit time. When sketching a new page, decide where the AI-quotable blocks should live, target them at ~150 words each, and surround them with shorter transition prose if needed. The goal is not 100% citability — even highly optimized pages have intros, conclusions, and link sentences that fall outside the sweet spot. The goal is to have at least 3 to 5 paragraphs in the ideal range per major page, each containing a complete self-contained fact-block. Pages with this structure get cited substantially more often than pages with one long perfect paragraph or twenty short snippets.

Frequently asked questions

Does this apply to non-English content? Roughly. The 134–167 range was measured on English corpora. Languages with denser tokens (German, Chinese) may have a slightly lower word count for the same character density. Treat the range as approximate for non-English content; the structural principle (self-contained, fact-dense blocks) translates directly.

Should I split existing long paragraphs immediately? Only on high-value pages — your homepage, top 5 landing pages, top 10 organic-traffic blog posts. Splitting a 500-page documentation site mechanically is not worth the effort. Focus on the pages where citation lift matters.

Does this change how I write headlines and bullet lists? No. Citability scoring only counts <p> content. Headings, list items, code blocks, and tables are evaluated separately by AI engines and have their own optimization patterns covered in our guide to llms.txt.